# Screening one-year tree-cover transitions in CONUS with annual satellite embeddings **Revised after review by an AI agent.** The fixed paired-native check remains unfinished; its empty counts are not robustness evidence. ## Abstract Tree-cover change maps can guide verification, but their scores need validation for the cover transition of interest. We retrospectively compared reference adjustments and classifiers for one-year transitions in the conterminous United States (CONUS) using AlphaEarth V1 mosaic and TESSERA v1, with initial classes supplied. The LCMAP tests included 109 gain cases and 337 loss cases (124 losses); 21 of 24 gains occurred in 2019–2020 and 3 in 2020–2021. A separate Annual NLCD Forest test used transferred thresholds. At nominal 10% earlier-negative false-positive calibration, reference/control corrections changed Youden’s J by -0.037 to +0.011, without consistent improvement. For year-matched references and control-mean adjustment, respectively, changed decisions numbered 3 and 9 for AlphaEarth gains, 19 and 6 for losses; TESSERA counts were 7 and 10 and 18 and 5. Only 1.5–9.2% of pooled test decisions changed, and rankings stayed similar (Spearman 0.947–0.996; post-review diagnostic). The fixed-reference score detected 3 of 24 AlphaEarth gains (AUROC 0.556) and 8 TESSERA gains. Cosine magnitude detected no gains in either product and ranked gain plots below non-tree negatives (AUROC 0.287/0.318). The direct supervised classifier had higher J in all 4 original conditions, with unequal label resources. Exploratory partition changes reversed all 8 correction contrasts, qualifying that comparison. An exploratory post-review unchanged-only comparison retained the gain-ranking inversion. Reference-centroid variation was smaller than loss-target changes, and the data do not identify timing, training supervision or physical growth as its cause. These small-perturbation, one-year tests do not establish performance for multi-year restoration trajectories or ecological recovery. ## 1. Introduction Monitoring vegetation recovery requires measurements that distinguish a change in land cover from recovery of the ecosystem being restored. The international standards of the Society for Ecological Restoration describe reference ecosystems and measurable indicators across ecological attributes; tree cover alone does not establish restoration success (Gann et al., 2019, Principles 3 and 5). Satellite observations can support monitoring over large areas, but a convenient numerical indicator still needs validation against the particular outcome it is intended to screen. Google Earth’s experimental Detect change tool now provides annual embedding-change layers at 10m and directs users to historical imagery to verify the type of change (Google, 2026, Detect change documentation). The Earth Engine catalogue also describes comparing annual embeddings by dot product or angle (Google and Google DeepMind, 2026, Satellite Embedding V1 catalogue). Heiman’s restoration application follows reference-similarity trajectories from restoration start across multiple years. These workflows make the choice of a change score relevant to selecting locations for more detailed imagery or field assessment; their effects on users were not measured here. Our saved-input comparison does not benchmark the Google Earth interface or its exact processing. Public annual satellite embeddings make this question timely. AlphaEarth Foundations and TESSERA provide vectors summarizing satellite observations that can be used in downstream mapping tasks (Brown et al., 2025; Feng et al., 2026). A recent restoration application measures cosine similarity to reference forest embeddings, including geographically selected references (Heiman, 2026). Such a reference score is attractive when detailed local outcome labels are scarce. Its interpretation depends on the reference, however, and changes in the target and reference vectors can both affect its value. Recomputing the reference every year and subtracting changes at independent control plots are plausible adjustments. Neither adjustment establishes that a transferred alert threshold detects the intended cover transition. We study a narrower, independently interpretable endpoint: transitions between dominant tree and non-tree vegetation cover. This is relevant to deciding whether a reference score is suitable for screening tree-cover gains and losses before more detailed assessment. The intended readers are restoration ecologists and land-cover monitoring teams deciding which scores and classifiers to validate before using them in assessment. The study measures screening errors, without measuring restoration success, decisions by practitioners, or savings in field work. The research question is: **in these CONUS one-year pairs, do year-matched references or separate control-plot mean changes improve transition classification over a fixed start-year reference and supervised classifiers?** The desired cover depends on the appropriate reference ecosystem; tree gain is not treated as inherently beneficial. Each case supplies its initial cover class. Gain screening concerns initially non-tree plots and loss screening initially tree-covered plots. Consequently, classifying the subsequent cover state is a strong and necessary comparator. The task classifies a transition with the initial class supplied; it does not infer both cover states from imagery. ## 2. Prior work and scope of the contribution AlphaEarth's evaluation includes LCMAP cover-change labels and a cosine-based angular anomaly score. Its reported evaluation uses validation-selected thresholds (Brown et al., 2025, Supplement S3.3.1 and S4.1). Our cosine-change comparator is a monotonic rescaling of that angular anomaly. Our question concerns thresholds calibrated on earlier negative examples outside the eventual test region. Differences from the published sample, labels, threshold rule and split preclude a direct comparison with its reported accuracy. Heiman (2026, Sections 2–5) evaluates cosine similarity to forest references for restoration polygons in the Brazilian Atlantic Forest. The reference labels use a longer stability history and the application differs from the US cover-transition endpoints studied here. We do not reproduce that ecological assessment or test its conclusions. We examine fixed, concurrent and control-adjusted reference scores using baseline-only reference selection and held-out cover labels. Change vector analysis already separates a temporal feature-difference vector from its Euclidean intensity; object-level methods also combine that intensity with correlation constraints (Wang et al., 2018, Sections 2.2.3–2.2.4). For unit vectors, the squared Euclidean difference is twice the cosine change magnitude. We do not propose that distinction or a new change-vector method. The present input comparisons concern the transferred operating points of published embeddings under supplied initial vegetation classes. Conditional change detection with pre-change semantic information is already established. Bernhard et al. (2023, Section 3 and Table 1) distinguish conditional tasks and a cross-modal setting using a prior semantic map and current imagery; their Section 5.4 evaluates that setting. Our supplied initial plot class similarly makes subsequent-state classification a legitimate comparator. We do not propose that task definition or reproduce MapFormer’s image segmentation experiment. Independent controls and baseline matching already have a substantial methodological history. Meroni et al. (2017, Sections 3–5) apply a before–after control-impact design to remote-sensing assessment of restoration interventions, with control selection and contextual checks. Our controls support a measurement comparison, not a causal estimate of an intervention. They are selected from cover labels and geography; they are not environmentally matched counterfactual ecosystems. Burns (2026), an author case study rather than a peer-reviewed comparison, already contrasts difference vectors, dot products, stacked annual embeddings and constructions retaining baseline context for landscape-change attribution. Sharma et al. (2026, Sections 2.3–2.4 and 4) combine annual AlphaEarth drift with spectral and thermal indicators and distinguish cover-transition evaluation from unconfirmed within-forest degradation. These applications reinforce the need to name the endpoint being validated. Singh et al. (2026, Section 2.1.2 and Table 1) already compare temporal reuse of baseline labels, stability filtering and semi-supervised updating under space-time validation. Our baseline-cohort scores and classifiers neither reproduce nor outperform that updating framework. Backman et al. (2026, Sections II–III) study supervised bitemporal woody clearing and transfer to regrowth using Sentinel-2 images; separate gain/loss evaluation is not itself a contribution here. Temporal variation of satellite embeddings and temporal degradation of learned classifiers are also known. Ball et al. (2026, Section 3.5, Table 5) transfer a tree-species classifier across years and discuss both changing representations and a disturbance affecting the landscape. Annual variation, control centring and temporal transfer are therefore not claimed as novel findings. The present contribution is a reproducible comparison of transferred operating points, explicit conditional-state comparators, and boundary checks across two published embedding products and two separately reported cover definitions. Searches are documented in sources.jsonl. An additional AlphaEarth/TESSERA temporal-transfer and domain-adaptation preprint was identified only at abstract level (Gbodjo and Ienco, 2026); its full methods were not reviewed and its results are not supporting evidence here. Failure to find a closer study within the searched range is not evidence that none exists. Established annual-change methods include LandTrendr, which segments spectral trajectories, and Continuous Change Detection and Classification (CCDC), which finds breakpoints with harmonic models (Google, 2026,algorithm documentation). They use temporal information beyond the single-year pairs evaluated here; neither was run as an additional comparator. The classic Malila (1980) CVA repository abstract describes spectral change vectors and forest harvesting/regrowth applications; only that abstract was obtained, so its empirical results are not evidence here. Wang et al. (2018) remains the checked full-text source for the vector/intensity distinction. Related work also includes Manikandan et al. (2026) on annual embedding change in a hyper-arid region, checked at abstract level only; full methods were unavailable. The CaGIS 2026 programme confirms Yao et al.’s interactive land-cover-change presentation; its extended abstract and reviewer-reported threshold were not obtained, so no numerical comparison is adopted. The official USGS abstract of Stehman et al. (2021), also abstract-only here, reports weak exact-year change agreement, improvement under a wider timing window and tree-cover loss as the strongest class-specific change. This contextual similarity does not show that timing explains the present sample’s results, and its operational-product accuracies are not directly comparable to this conditional, enriched sample. ## 3. Methods ### 3.1 Labels, sampling and endpoints The main labels are the public-domain Land Change Monitoring, Assessment, and Projection (LCMAP) component of the AlphaEarth supplemental evaluation release (Google DeepMind, 2025), derived from the USGS reference data (Pengra et al., 2023). We use its human-interpreted dominant-cover classes Trees, Grass/forb/herb and Shrubs and adjacent pairs spanning 2017–2021. LCMAP Trees includes standing dead trees. Dominant Trees is not a verified native forest reference ecosystem or a forest land-use classification. The source benchmark selects one adjacent pair per plot and enriches changes; the resulting sample does not estimate national transition prevalence. We reserve all eligible Trees-to-Trees plots in the first pair for the reference/control pool and exclude them from the targets. Selection uses only the baseline pair and never requires future stability. The baseline pool contains 116 plots: 58 reference plots and 58 separate control plots before region exclusions, allocated by a fixed hash. The annual valid counts outside each region are reported in Appendix E. Here independent controls means disjoint plot membership, not randomized treatment assignment or statistical independence of landscapes. The primary allocation uses seed 14014; four further allocations are sensitivity analyses on the same pool. All eligible tree gains and losses and later stable Trees pairs are retained. Non-tree negative examples are capped by class and year using a fixed hash. A separate baseline-selected non-tree control cohort supports prespecified gain-screening comparisons. Exact counts, deterministic rules and exclusions are generated by scripts/prepare.py and scripts/prepare_controls.py and specified in protocol/plan-v1.md and plan-v2.md. The separate Annual National Land Cover Database (NLCD) reference-data extension uses Collection 1.0 annual interpreted cover labels, without fitting or recalibrating to them (USGS, 2025). Its Forest class specifies at least 20% tree vegetation cover with trees greater than 5 metres tall; Shrubland includes young trees below 5 metres. We retain original labels and use a computational class crosswalk solely to apply the scores and classifiers. This endpoint differs from LCMAP dominant Trees. The extension excludes nearby plots using coordinates, not ownership or personal information. Overlapping years and years beyond the original benchmark are reported separately; they are not pooled into a common ecological outcome or attributed solely to elapsed time. Primary labels and the subset agreeing with the alternate cover label are both reported. ### 3.2 Representations and spatial support We use published AlphaEarth Foundations V1 annual embeddings and TESSERA v1 embeddings, not newly trained encoders. AlphaEarth's primary source is the unofficial geographic mosaic documented by tge-labs (2026); TESSERA is pinned to the v1 public store. AlphaEarth signed-square quantization is decoded before normalization following the official documentation (Google and Google DeepMind, 2026). Every pixel vector is normalized to unit length. Centre pixels are primary. A support sensitivity normalizes the mean of the nine unit vectors in a 3-by-3 patch and requires all nine pixels to be valid. The mosaic has approximately 0.00008983 degrees per pixel in each axis; the retrieved TESSERA grid has 10-metre pixels in its projected CRS. The geographic mosaic's footprint varies with latitude and is not an exact 30-by-30-metre plot footprint. Inputs were retrieved on 2026-10-25 (UTC). Source URLs, version metadata, coordinates, transforms, licences, raw arrays and hashes are bundled. AlphaEarth distinguishes the represented valid period from the supporting observations (Brown et al., 2025, Section 2); exact support periods for these mosaics were not independently verified. TESSERA documents annual observation sampling (Feng et al., 2026, Appendix A.4). Neither encoder was trained prospectively in this study. Brown et al. (2025, S1.9 and Tables S1–S2) document NLCD weak supervision: their evaluated v2.0 model uses the 2019 release outside 2021, with nearest mapped-year assignment and loss weight 0.5. The official 2019-release catalogue contains multiple map epochs, including 2016 and 2019; this does not establish that every 2017–2020 sample was trained against one static 2019 map (Google/USGS, 2026). Their S8.3 and the official catalogue describe production v2.1, adding USDA Cropland Data Layer (CDL) targets through 2023 and lowering NLCD/CDL weights to 0.25. The unofficial V1-labelled mosaic does not verify its exact encoder lineage. This is a potential dependence, not an identified cause of our scores. The published TESSERA method instead uses Barlow Twins self-supervision of Sentinel-1/2 sequences (Feng et al., 2026). A separately fixed stratified sample compares the AlphaEarth mosaic with official native-grid COG measurements at the same requested coordinates. It reports vector agreement and pixel-centre displacement, without substituting nearby valid pixels. This is an input co-location diagnostic, not a full native-grid classification evaluation. Agreement cannot rule out small changes to threshold decisions. ### 3.3 Geographic and temporal transfer Five regions are formed by coordinate-only KMeans in a contiguous-US projection. For each held-out region, earlier pairs beginning in 2017 or 2018 outside that region supply fitting and calibration examples. Later pairs beginning in 2019 or 2020 inside it supply test examples. A deterministic hash of five-degree geographic cells reserves calibration cells separately from model-fitting cells. The 2019 endpoint of the earlier pairs overlaps the first later pair’s starting year in other regions. Later target plots are never used in fitting. The same target partitions are offered to all methods, although their actual use of earlier labels differs. This is not an annotation-budget-matched comparison. Reference and control plots inside the held-out region are excluded. Contemporary embeddings of baseline-selected plots are permitted. Encoders and interpreted labels were produced after some represented years; the design tests retrospective downstream transfer, not a reconstruction of information available in 2019 or a holdout from encoder pretraining. Primary estimates use target pairs available in both products; model-specific availability is reported when it differs. Missing measurements do not cause replacement sampling. ### 3.4 Change scores, classifiers and calibration Let a and b be target unit vectors before and after a pair, and r_a and r_b normalized reference means in the corresponding years outside the held-out region. The fixed start-year reference score is (b−a)·r_a, the end-fixed score (b−a)·r_b, and the pooled-fixed score (b−a)·normalize(r_a+r_b). The concurrent score is b·r_b−a·r_a. Algebraically, its difference from the fixed start-year reference score is b·(r_b−r_a), so this comparison mixes reference motion with the target endpoint. Tree-control adjustment subtracts the independent controls' mean fixed start-year reference change. Standardized concurrent scores first centre and scale each year's reference similarity using independent controls, then take the difference. Vanishing control standard deviations are marked unavailable. Directed scores are negated for loss; cosine change magnitude, 1−a·b, is positive for both gain and loss. Subsequent-reference similarity uses only b·r_b, with the direction appropriate to each task. Learned comparisons include a random forest fitted to both cover states, its probability difference and subsequent-state probability, a subsequent-state logistic classifier, and a direct gradient-boosted transition classifier. All fitting uses only the earlier fitting partition. Prespecified hyperparameters are in the protocol and code; no test-based tuning or selection of a winning correction is performed. Additional pre-analysis comparisons use nearest available references, independent non-tree controls, initial-class control matching and a reference frozen to 2017. Geographic proximity is not assumed to establish ecological matching. For every method, task and region, the threshold is the higher-interpolated ninetieth percentile of earlier negative calibration scores. A flag requires strict exceedance. At least ten negative calibration cases and eight valid reference/control plots are required where applicable. Finite samples and ties can make calibration more conservative than the nominal ten-percent false-positive rate. This rule does not guarantee the same rate after transfer. The NLCD extension reuses the original fitting and calibration data and explicitly checks equality of thresholds. ### 3.5 Additional checks and diagnostics after review Original pre-analysis amendments and post-result exploratory checks are retained in the appendices and their committed protocols. They test control selection, calibration levels, reference membership/access, input periods, coordinate/class baselines, matched direct-learner inputs, partition sensitivity and native-input cases; none replaces the primary sample or threshold. Detailed settings and all failed/untested conditions remain referable. After REVIEW-1, a separate fixed plan adds descriptive diagnostics from saved data: original fold/year-valid reference centroids and individual baseline reference/control changes; pooled and year-specific Spearman score similarity and changed decisions; sample counts by pair/region; and magnitude medians/AUROC against both the original negatives and unchanged-only negatives. These are exploratory post-review analyses. No fitting, threshold, exclusion or prediction changes. Reference-only multi-year summaries do not evaluate multi-year target screening. ### 3.6 Outcomes, uncertainty and process We report sensitivity (producer’s accuracy), false-positive rate and their difference, Youden's J, with counts. The primary contrasts are concurrent versus fixed start-year reference and tree-control-adjusted versus fixed start-year reference, separately for both tasks and products. J summarizes sensitivity minus false-positive rate at the alert threshold, without specifying social costs. False-positive rate is not commission error; enriched sampling prevents interpreting the observed sample precision as population user’s accuracy. AUROC is a secondary descriptive diagnostic for pooled held-out scores; it does not isolate within-year or within-region ranking. We retain all planned comparisons, unavailable conditions, threshold ties, unchanged flags and reference-anchor disagreements. Two thousand paired geographic-cell bootstrap resamples give descriptive percentile intervals conditional on fitted scores, observed targets and the fixed reference pool. Resamples without both positive and negative cases are omitted; valid-draw counts are retained. These intervals omit training, encoder and independent future-year variability. At observed error floors and ceilings they can collapse, without establishing perfect population performance. Reference allocations are sensitivity analyses, not independent replications. Constant control subtraction preserves rankings within a fixed year and region by algebra; only transferred decisions can demonstrate empirical improvement. Plans were committed before measurements or analyses to which they apply. Technical failures, timing-related acquisition changes and departures are retained in protocol/deviations.md and experiment records. Analyses fixed after the primary results are identified as exploratory. Code, public selected inputs and a network-free reproduction entry point are supplied; the verification procedure is VERIFY.md. Local Git, timestamps and hashes provide inspectable records, not proof of independent chronology or reproduction. ## 4. Results ### 4.1 What the test sample supports The original test has 24 gain positives and 124 loss positives. Gains are concentrated in 2019–2020: 21 cases, versus 3 in 2020–2021. The table gives every year-pair and region denominator; anonymous region identifiers are split indices, not regional rankings. The small second-pair and regional samples cannot establish broad temporal/geographic gain transfer. The separate later Forest test has few gains too. Initial cover class is supplied in every case. | Grouping | Pair / anonymous region | Gain positives | Gain negatives | Loss positives | Loss negatives | Gain-positive cells | | --- | --- | --- | --- | --- | --- | --- | | Year pair | 2019–2020 | 21 | 46 | 68 | 117 | 12 | | Year pair | 2020–2021 | 3 | 39 | 56 | 96 | 3 | | Held-out region | 0 | 5 | 50 | 40 | 46 | 3 | | Held-out region | 1 | 6 | 22 | 25 | 30 | 3 | | Held-out region | 2 | 3 | 4 | 5 | 25 | 2 | | Held-out region | 3 | 8 | 8 | 45 | 65 | 4 | | Held-out region | 4 | 2 | 1 | 9 | 47 | 2 | ### 4.2 Reference adjustments mostly preserve the original score Table 1 gives absolute discrimination as well as sensitivity and false-positive rate. AlphaEarth gain AUROC is close to chance under the fixed-reference direction; corrections do not make it a strong gain ranking. Table 2 retains every primary correction contrast and decision count. The pooled Spearman range is 0.947–0.996, with alert status changed in 1.5–9.2% of cases. These similarity measurements were added after review. For the global control mean, subtraction is constant within region/year by construction (maximum measured SD 7.2e-18). Near-identical scores and few changed decisions limit the strength of the correction comparison; a lack of consistent benefit is not evidence against correction under substantial harmful drift. | Representation | Task | Change score | Detected / positive | False alerts / negative | Sensitivity | FPR | J | AUROC | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Fixed start-year reference | 3/24 | 9/85 | 0.125 | 0.106 | 0.019 | 0.556 | | AlphaEarth V1 mosaic | gain | Year-matched reference | 3/24 | 12/85 | 0.125 | 0.141 | -0.016 | 0.510 | | AlphaEarth V1 mosaic | gain | Control-mean adjustment | 5/24 | 16/85 | 0.208 | 0.188 | 0.020 | 0.533 | | AlphaEarth V1 mosaic | loss | Fixed start-year reference | 64/124 | 18/213 | 0.516 | 0.085 | 0.432 | 0.696 | | AlphaEarth V1 mosaic | loss | Year-matched reference | 60/124 | 19/213 | 0.484 | 0.089 | 0.395 | 0.694 | | AlphaEarth V1 mosaic | loss | Control-mean adjustment | 62/124 | 22/213 | 0.500 | 0.103 | 0.397 | 0.695 | | TESSERA v1 | gain | Fixed start-year reference | 8/24 | 11/85 | 0.333 | 0.129 | 0.204 | 0.641 | | TESSERA v1 | gain | Year-matched reference | 8/24 | 12/85 | 0.333 | 0.141 | 0.192 | 0.613 | | TESSERA v1 | gain | Control-mean adjustment | 6/24 | 3/85 | 0.250 | 0.035 | 0.215 | 0.661 | | TESSERA v1 | loss | Fixed start-year reference | 93/124 | 19/213 | 0.750 | 0.089 | 0.661 | 0.862 | | TESSERA v1 | loss | Year-matched reference | 99/124 | 27/213 | 0.798 | 0.127 | 0.672 | 0.858 | | TESSERA v1 | loss | Control-mean adjustment | 92/124 | 15/213 | 0.742 | 0.070 | 0.672 | 0.857 | Table 2. Correction minus fixed-reference J, conditional paired intervals and score similarity. Most decisions are unchanged. These intervals are not equivalence tests and omit fitting/reference uncertainty. Pooled similarity does not mean each year is equally similar: year-specific correlations and alert changes are retained, with changed shares up to 21.4% in a smaller year-specific condition. | Representation | Task | Adjustment versus fixed reference | J difference | Conditional 95% interval | Changed alerts / cases | Pooled rank correlation | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Year-matched reference | -0.035 | -0.079 to 0.000 | 3/109 | 0.947 | | AlphaEarth V1 mosaic | gain | Control-mean adjustment | 0.001 | -0.091 to 0.117 | 9/109 | 0.979 | | AlphaEarth V1 mosaic | loss | Year-matched reference | -0.037 | -0.100 to 0.014 | 19/337 | 0.960 | | AlphaEarth V1 mosaic | loss | Control-mean adjustment | -0.035 | -0.064 to -0.011 | 6/337 | 0.982 | | TESSERA v1 | gain | Year-matched reference | -0.012 | -0.074 to 0.051 | 7/109 | 0.957 | | TESSERA v1 | gain | Control-mean adjustment | 0.011 | -0.138 to 0.124 | 10/109 | 0.996 | | TESSERA v1 | loss | Year-matched reference | 0.011 | -0.039 to 0.087 | 18/337 | 0.964 | | TESSERA v1 | loss | Control-mean adjustment | 0.011 | -0.011 to 0.035 | 5/337 | 0.991 | ### 4.3 How much annual variation was available to correct? The post-review diagnostic reconstructs the original region-excluded, per-year-valid reference centroids. Values below are cosine dissimilarity (one minus the dot product). Centroid/control ranges span the five held-out regions; target medians pool original eligible paired cases in that year, with counts. Reference/control plots were Trees at baseline, not verified stable ecosystems in later years. Earlier target samples supply context, not extra test evidence. | Representation | Year pair | Reference-centroid dissimilarity range | Control-plot median range | Gain-target median | Loss-target median | | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 2017–2018 | 0.0071–0.0116 | 0.0248–0.0279 | 0.050 (n=6) | 0.153 (n=59) | | AlphaEarth V1 mosaic | 2018–2019 | 0.0063–0.0107 | 0.0209–0.0236 | 0.033 (n=78) | 0.115 (n=59) | | AlphaEarth V1 mosaic | 2019–2020 | 0.0026–0.0057 | 0.0192–0.0216 | 0.033 (n=21) | 0.191 (n=68) | | AlphaEarth V1 mosaic | 2020–2021 | 0.0035–0.0064 | 0.0205–0.0234 | 0.027 (n=3) | 0.243 (n=56) | | TESSERA v1 | 2017–2018 | 0.0601–0.0651 | 0.1419–0.1644 | 0.144 (n=6) | 0.360 (n=59) | | TESSERA v1 | 2018–2019 | 0.0099–0.0126 | 0.0942–0.1028 | 0.095 (n=78) | 0.253 (n=59) | | TESSERA v1 | 2019–2020 | 0.0083–0.0105 | 0.0903–0.0963 | 0.090 (n=21) | 0.310 (n=68) | | TESSERA v1 | 2020–2021 | 0.0193–0.0270 | 0.0936–0.1084 | 0.114 (n=3) | 0.401 (n=56) | AlphaEarth centroid variation is small relative to the loss-target magnitude in these pairs. TESSERA's 2017–2018 centroid variation is larger than its next two adjacent pairs and lies in the earlier fitting/calibration period. Thus the products present different opportunities for an adjustment; neither variation identifies representation drift separately from physical change. Appendix E retains every fold/year count, individual reference/control variation, same-membership centroid check and descriptive2017-to-2023 comparisons. Multi-year target screening was not tested. ### 4.4 Reference direction, magnitude and supervised classification differ The direct transition classifier has higher J point estimates than the fixed-reference direction in all four original product/task conditions. It uses more earlier labels, so this is not equal-budget superiority. Given initial cover, classifying the subsequent state is also a necessary comparator. For AlphaEarth losses, magnitude is much stronger than the reference direction: AUROC 0.968 versus 0.696, with sensitivities 0.903 versus 0.516. The asymmetry is observed on these labelled pairs, not a new universal property of vegetation change. | Representation | Task | Score / classifier | Detected / positive | False alerts / negative | Sensitivity | FPR | J | AUROC | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Cosine change magnitude | 0/24 | 7/85 | 0.000 | 0.082 | -0.082 | 0.287 | | AlphaEarth V1 mosaic | gain | Direct gradient boosting | 12/24 | 6/85 | 0.500 | 0.071 | 0.429 | 0.866 | | AlphaEarth V1 mosaic | gain | State-probability difference (forest) | 5/24 | 5/85 | 0.208 | 0.059 | 0.150 | 0.663 | | AlphaEarth V1 mosaic | gain | Subsequent-reference similarity | 5/24 | 5/85 | 0.208 | 0.059 | 0.150 | 0.875 | | AlphaEarth V1 mosaic | gain | Subsequent-state logistic | 12/24 | 20/85 | 0.500 | 0.235 | 0.265 | 0.759 | | AlphaEarth V1 mosaic | gain | Subsequent-state forest | 9/24 | 9/85 | 0.375 | 0.106 | 0.269 | 0.846 | | AlphaEarth V1 mosaic | loss | Cosine change magnitude | 112/124 | 27/213 | 0.903 | 0.127 | 0.776 | 0.968 | | AlphaEarth V1 mosaic | loss | Direct gradient boosting | 111/124 | 33/213 | 0.895 | 0.155 | 0.740 | 0.957 | | AlphaEarth V1 mosaic | loss | State-probability difference (forest) | 105/124 | 8/213 | 0.847 | 0.038 | 0.809 | 0.951 | | AlphaEarth V1 mosaic | loss | Subsequent-reference similarity | 54/124 | 51/213 | 0.435 | 0.239 | 0.196 | 0.754 | | AlphaEarth V1 mosaic | loss | Subsequent-state logistic | 100/124 | 32/213 | 0.806 | 0.150 | 0.656 | 0.890 | | AlphaEarth V1 mosaic | loss | Subsequent-state forest | 113/124 | 49/213 | 0.911 | 0.230 | 0.681 | 0.929 | | TESSERA v1 | gain | Cosine change magnitude | 0/24 | 8/85 | 0.000 | 0.094 | -0.094 | 0.318 | | TESSERA v1 | gain | Direct gradient boosting | 14/24 | 8/85 | 0.583 | 0.094 | 0.489 | 0.874 | | TESSERA v1 | gain | State-probability difference (forest) | 3/24 | 5/85 | 0.125 | 0.059 | 0.066 | 0.737 | | TESSERA v1 | gain | Subsequent-reference similarity | 11/24 | 7/85 | 0.458 | 0.082 | 0.376 | 0.907 | | TESSERA v1 | gain | Subsequent-state logistic | 14/24 | 10/85 | 0.583 | 0.118 | 0.466 | 0.825 | | TESSERA v1 | gain | Subsequent-state forest | 14/24 | 5/85 | 0.583 | 0.059 | 0.525 | 0.869 | | TESSERA v1 | loss | Cosine change magnitude | 115/124 | 40/213 | 0.927 | 0.188 | 0.740 | 0.956 | | TESSERA v1 | loss | Direct gradient boosting | 114/124 | 30/213 | 0.919 | 0.141 | 0.779 | 0.959 | | TESSERA v1 | loss | State-probability difference (forest) | 102/124 | 21/213 | 0.823 | 0.099 | 0.724 | 0.929 | | TESSERA v1 | loss | Subsequent-reference similarity | 66/124 | 43/213 | 0.532 | 0.202 | 0.330 | 0.797 | | TESSERA v1 | loss | Subsequent-state logistic | 109/124 | 51/213 | 0.879 | 0.239 | 0.640 | 0.938 | | TESSERA v1 | loss | Subsequent-state forest | 119/124 | 45/213 | 0.960 | 0.211 | 0.748 | 0.950 | ![Screening sensitivity and false-positive rates](figures/operating-points.png) For gains, magnitude AUROC is 0.287 for AlphaEarth and 0.318 for TESSERA on the original negatives. The inversion remains when the comparison uses unchanged non-tree plots only: pooled AUROCs are 0.339 and 0.352. The descriptive year-pair table shows smaller median embedding changes for gain plots than unchanged grass/shrub plots in both products and both pairs, with only three gains in the second pair. Unchanged means the same interpreted cover class, not verified absence of within-class structural, seasonal or management change. These unchanged-only ranking diagnostics were added after review; no threshold or prediction was changed. The earlier unchanged-negative calibration sensitivity also retains its zero gain flags (Appendix F). | Representation | Year pair | Gains | Unchanged negatives | Median gain magnitude | Median unchanged magnitude | AUROC versus unchanged | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 2019–2020 | 21 | 32 | 0.033 | 0.042 | 0.336 | | AlphaEarth V1 mosaic | 2020–2021 | 3 | 32 | 0.027 | 0.046 | 0.260 | | TESSERA v1 | 2019–2020 | 21 | 32 | 0.090 | 0.119 | 0.339 | | TESSERA v1 | 2020–2021 | 3 | 32 | 0.114 | 0.121 | 0.406 | ### 4.5 Separate Forest-class test The NLCD extension contains 523 selected pairs from 488 plots. Paired centre availability is 522/523. Its primary denominators are 2022–2023 gain: 7 positives/115 negatives; 2022–2023 loss: 26 positives/100 negatives; 2020–2021 gain: 24 positives/121 negatives; 2020–2021 loss: 29 positives/100 negatives. Pairs ending in 2020–2021 pairs overlap the original calendar range; 2022–2023 pairs extend it. Neither endpoint differences nor horizon differences can be attributed solely to time. All 620 original-method threshold comparisons match the LCMAP calibration, with the added initial/cohort/coordinate/class-rate thresholds checked separately. No NLCD outcome enters fitting or calibration. The primary correction contrasts and counts are below. Intervals remain conditional descriptive block intervals; the small positive counts limit their interpretation. The later gain positives occupy 4 five-degree cells; the corresponding correction intervals retain 1982 valid draws after omitting resamples without both classes. These support counts were added during self-check; the original bootstrap draws and intervals are unchanged. | Representation | Pair end years | Task | Positive / negative | Start-fixed J | Concurrent J difference [interval] | Tree-control J difference [interval] | Changed flags concurrent/control | | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 2022–2023 | gain | 7/115 | -0.191 | +0.000 [-0.059, 0.050] | -0.035 [-0.067, -0.008] | 6/4 of 122 | | AlphaEarth V1 mosaic | 2022–2023 | loss | 26/100 | 0.537 | -0.018 [-0.090, 0.039] | -0.012 [-0.081, 0.055] | 3/6 of 126 | | TESSERA v1 | 2022–2023 | gain | 7/115 | -0.083 | -0.017 [-0.078, 0.043] | -0.065 [-0.937, 0.150] | 8/20 of 122 | | TESSERA v1 | 2022–2023 | loss | 26/100 | 0.615 | -0.115 [-0.350, 0.021] | +0.055 [-0.184, 0.202] | 17/22 of 126 | | AlphaEarth V1 mosaic | 2020–2021 | gain | 24/121 | 0.117 | -0.158 [-0.452, 0.008] | -0.066 [-0.117, -0.022] | 9/8 of 145 | | AlphaEarth V1 mosaic | 2020–2021 | loss | 29/100 | 0.220 | -0.064 [-0.142, -0.010] | +0.004 [-0.058, 0.093] | 4/4 of 129 | | TESSERA v1 | 2020–2021 | gain | 24/121 | -0.048 | -0.025 [-0.170, 0.075] | -0.018 [-0.239, 0.131] | 11/16 of 145 | | TESSERA v1 | 2020–2021 | loss | 29/100 | 0.357 | -0.050 [-0.113, 0.034] | -0.034 [-0.113, 0.033] | 11/5 of 129 | The complete scores and classifiers, magnitude-score counts, initial-versus-subsequent comparisons and primary/alternate-agreeing subset are retained in Appendix B and the CSV outputs. Agreement means identical primary and alternate interpreted cover classes at both dates. It is not a verified confidence level, absence of within-class change or ecological recovery. Fixed-reference J is negative in 3 of 4 gain conditions. The separate reference sample and its unchanged calibration are useful boundary checks, but AlphaEarth's NLCD training targets prevent calling its Forest endpoint independent of training-target information. The map targets and interpreted reference annotations are different products. No ecological restoration outcome is validated. ### 4.6 Exploratory input and calibration boundaries The matched direct-learner input comparison retains before-only gain discrimination, while difference-only versus concatenated inputs behave differently for gains and losses. This does not establish forecasting or an information source; represented dates, training targets and retrospective encoder production matter. Full conditions, counts and conditional intervals remain in Appendix D. All 8 correction contrasts take both signs across the 10 fixed earlier fitting/calibration partitions. The direct TESSERA loss advantage holds in 6 of those partitions. The AlphaEarth centre-magnitude exception detects 1 gain with 17 false alerts in an alternative split, so the original zero count is not a universal floor. These exploratory conditions share targets and earlier data; they do not estimate independent training variance. Complete tables remain in Appendix F. Coordinate, initial-class, reference-access and native-input diagnostics have also moved there. The fixed paired-native test remains untested, not a completed robustness result. ## 5. Discussion The absolute scores matter before choosing a correction. In this labelled sample, a reference-forest direction ranks AlphaEarth gains near chance, and annual cosine magnitude ranks gains below the unchanged non-tree negatives. A high-change map therefore does not automatically screen these tree gains. Larger changes within an unchanged cover class may still describe useful phenomena that these labels do not measure. The losses have much larger embedding changes and are better ranked by magnitude. These are endpoint-specific observations on predominantly one gain year pair; they do not show that representations lack gain information, that gains are gradual, or that the labels are wrong. The correction null is a limited comparison. Small centroid motion, highly correlated scores and mostly unchanged alerts make it unsurprising that changes in J are small. Control subtraction also preserves rankings within a fixed region/year by algebra. The measured transferred decisions can still change, including with earlier calibration choices; their signs are not guaranteed by that algebra. TESSERA's larger 2017–2018 variation shows why calibration-period variation should be reported. We have not tested correcting large harmful representation shifts, ecologically matched local controls or multi-year restoration-start trajectories. Possible explanations for the gain/loss pattern include cover-boundary timing, plot/pixel mismatch, stable landscape associations and training-target dependence. None is isolated here. The known AlphaEarth supervision described in §3.2 could contribute to dated-vector association or temporal smoothing. The similar TESSERA pattern shows that AlphaEarth's particular supervision is not necessary for the qualitative observation, without identifying a causal mechanism. The two products are informative contrasts, not matched independent replications of a training design. Operational annual-change mapping already faces timing and gain/loss difficulties; the abstract-level Stehman comparison is consistent context, not proof that timing explains these observations. LandTrendr and CCDC use longer temporal information than a single annual pair. Their names identify established alternatives, without a performance comparison in this study. Similarly, the current Google Earth workflow motivates endpoint validation but was not benchmarked: its exact processing and thresholds, unlike our saved mosaic inputs, were not evaluated. For restoration ecologists and land-cover monitoring teams, the supported decision is to validate the intended cover transition and alert threshold before using embedding change or reference similarity to select locations for imagery or field assessment. Report absolute sensitivity, false-positive rate, score ranking and year-specific counts, alongside any correction difference. Given an initial class, include subsequent-state and initial-date-only comparisons. These suggestions concern evaluation; benefits to users and ecological restoration success were not measured. ## 6. Limitations Only one-year intervals in CONUS were classified. 21 of 24 LCMAP gains belong to 2019–2020 and only 3 to 2020–2021, with small negative counts in some held-out regions. The gain conclusions chiefly describe the first pair. Later Forest gains are limited too; exact counts appear in §4.5. Descriptive multi-year reference drift is not a multi-year target-classification experiment. Enriched/capped samples do not estimate national prevalence, deployed precision or commission error. The reference/control construction pools a small baseline-labelled sample across CONUS outside each held-out region. It is a broad reference direction, not a verified local reference ecosystem. Later cohort stability is unknown; isolated auxiliary labels do not establish full histories. Conclusions about corrections apply to that pooled construction, not arbitrary restoration-reference designs. AlphaEarth uses land-cover weak supervision (§3.2); independence of its Forest endpoint from training-target information is not established. Product-family V1 does not verify the unofficial mosaic's exact encoder lineage, and public production-model changes further limit an exact training-mechanism assignment. Supporting-observation periods were not independently verified. Encoders/labels were produced retrospectively, and a boundary year overlaps the earlier/later periods. Before-only association is not forecasting or proof of leakage. The native-grid paired check remains unfinished; single-date substitutions cannot validate full native performance. Bootstrap intervals condition on fitted scores and the observed reference pool, omit training/encoder/future-year variance, and can collapse at error floors. Extra partitions and post-review diagnostics remain exploratory. The process retains a brief concurrency-limit breach, acquisition/reporting failures and lost arrays; those disclosures and local Git/hashes do not establish independent chronology. The planned review was by an AI agent; its reported rerun on the same machine is not an independent reproduction. ## 7. Reproduction and data attribution VERIFY.md describes the offline entry point. The recorded end-to-end self-rerun returned PASS. It covers the current numerical code. This is a rerun by the authoring AI session, not independent reproduction. Details and runtime are in review/self-rerun-4.json. Selected raw measurements, public labels, fixed plans, code, predictions and failure records are bundled. No live data download or AI call is required to recompute the scientific estimates. The authoring session’s sole other-model API generation was the permitted topic-candidate critique; it was not a paper review, research assistant or experimental model measurement. The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind. AlphaEarth data are CC BY 4.0; TESSERA v1 data are CC0; analysed LCMAP and Annual NLCD reference data are public domain. The full input archive contains unused components with separate terms. Source records distinguish relevant full-text checks from abstract-only search records. The human supplied one planned review by an AI agent (review/review-1.md). Its reported rerun on the same machine and its model/settings are human-provided reports, not directly observed author measurements or independent reproduction. Author revisions and decisions are in review/response-1.md. ## References Gann, G. D., et al. (2019). [International principles and standards for the practice of ecological restoration. Second edition](https://doi.org/10.1111/rec.13035). Restoration Ecology. Meroni, M., et al. (2017). [Remote sensing monitoring of land restoration interventions in semi-arid environments with a before–after control-impact statistical design](https://doi.org/10.1016/j.jag.2017.02.016). International Journal of Applied Earth Observation and Geoinformation, 59, 42–52. Brown, C. F., et al. (2025). [AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data](https://arxiv.org/abs/2507.22291v2). arXiv:2507.22291v2. Feng, Z., et al. (2026). [TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis](https://arxiv.org/abs/2506.20380v7). arXiv:2506.20380v7; first preprint 2025. Heiman, A. (2026). [Characterizing Brazilian Atlantic Forest Restoration Outcomes with Geospatial AlphaEarth Embeddings](https://arxiv.org/abs/2605.05547v1). arXiv:2605.05547v1, ICLR ML4RS workshop preprint. Ball, J. G. C., et al. (2026). [Geospatial foundation models enable data-efficient tree species mapping in temperate mountain forests](https://doi.org/10.1016/j.srs.2026.100466). Science of Remote Sensing, article 100466. Pengra, B. W., et al. (2023). [LCMAP CONUS Reference Data Product 1984–2021 land cover, land use and change process attributes](https://doi.org/10.5066/P933Z1TK). U.S. Geological Survey data release. Google DeepMind (2025). [AlphaEarth Foundations Supplemental Evaluation Datasets, v1.0.0](https://doi.org/10.5281/zenodo.16585402). Zenodo data release. U.S. Geological Survey (2025). [Annual National Land Cover Database (NLCD) Collection 1.0 Reference Data Product](https://doi.org/10.5066/P13EDMAF). U.S. Geological Survey data release; official MRLC archive last modified 29 April 2026. Google and Google DeepMind (2026). [AlphaEarth Foundations satellite embeddings on Google Cloud Storage](https://developers.google.com/earth-engine/guides/aef_on_gcs_readme). Official data access and dequantization documentation, checked 25 October 2026. TESSERA project (2026). [TESSERA data and GeoTessera documentation](https://github.com/ucam-eo/tessera). Official project data documentation, checked 25 October 2026. Burns, M. (2026). [Rethinking Change Detection and Attribution: How You Compare Satellite Embeddings Matters](https://medium.com/google-earth/rethinking-change-detection-and-attribution-how-you-compare-satellite-embeddings-matters-858f17f577d7). Author case study, Google Earth and Earth Engine, 13 April. Sharma, A., et al. (2026). [Label-Free Detection of Hidden Forest Degradation Using Sentinel-2, Landsat Thermal Stress and AlphaEarth Embeddings](https://www.preprints.org/manuscript/202610.0842). Preprints.org, version1, not peer reviewed. Gbodjo, Y. J. E., and Ienco, D. (2026). [Assessing the generalization of geospatial embeddings for temporal transfer in land cover classification](https://doi.org/10.2139/ssrn.6890917). SSRN preprint; abstract-only search record, full methods not reviewed. Bernhard, M., Strauß, N., and Schubert, M. (2023). [MapFormer: Boosting Change Detection by Using Pre-change Information](https://arxiv.org/abs/2303.17859v4). ICCV2023; arXiv2303.17859v4. Singh, G., Slingsby, J. A., Robinson, T. B., and Moncrieff, G. (2026). [Reducing the labeling burden in time-series mapping using Common Ground: a semi-automated approach to tracking changes in land cover and species over time](https://arxiv.org/pdf/2602.04373). arXiv:2602.04373v1. Backman, K., Wood, J., and Roff, A. (2026). [Learning Woody Clearing with Loss Alignment for Zero-Shot Regrowth and Woody Segmentation](https://arxiv.org/html/2608.26489v1). arXiv:2608.26489v1, 27 August 2026. Wang, Z., Liu, Y., Ren, Y., and Ma, H. (2018). [Object-Level Double Constrained Method for Land Cover Change Detection](https://doi.org/10.3390/s19010079). Sensors 19(1), 79; published 26 December 2018. tge-labs (2026). [AEF Mosaic README](https://data.source.coop/tge-labs/aef-mosaic/README.md). Source Cooperative account documentation for the unofficial public mosaic, checked 25 October 2026. Google (2026). [Detect change (Experimental)](https://developers.google.com/maps/documentation/earth/detect-change). Documentation last updated2026-10-23; relevant full documentation checked. Google and Google DeepMind (2026). [Satellite Embedding V1 Earth Engine catalogue](https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL). Current catalogue checked2026-10-25; relevant full documentation checked. Stehman, S.V., Pengra, B., Horton, J., Wellington, D.F. (2021). [Validation of the U.S. Geological Survey’s LCMAP collection1.0 annual land cover products1985–2017](https://www.usgs.gov/publications/validation-us-geological-surveys-land-change-monitoring-assessment-and-projection). 2021; DOI10.1016/j.rse.2021.112646; abstract only. Yao, X., Wu, K., Bhatt, R., Hall, B., Kang, Y. (2026). [Cartographer in the Loop: Interactive Land Cover Change Detection Using Geospatial Foundation Model Embeddings](https://cartogis.org/conferences/cagis2026/program-old/). CaGIS2026 programme record; programme record only. Google (2026). [LandTrendr algorithm documentation](https://developers.google.com/earth-engine/apidocs/ee-algorithms-temporalsegmentation-landtrendr). Documentation checked2026-10-25; relevant full documentation checked. Malila, W.A. (1980). [Change Vector Analysis: An Approach for Detecting Forest Changes with Landsat](https://docs.lib.purdue.edu/lars_symp/385/). 1980; abstract only. Manikandan, K.P., et al. (2026). [Foundation-Model Embeddings for Land-Cover Mapping and Annual Change Detection in a Hyper-Arid Region: A Case Study of Saudi Arabia (2017–2024)](https://doi.org/10.3390/rs18183163). Remote Sensing18(18),3163; abstract only, full methods not checked. Google (2026). [CCDC algorithm documentation](https://developers.google.com/earth-engine/apidocs/ee-algorithms-temporalsegmentation-ccdc). Official documentation; full relevant description checked25October2026. Google/USGS (2026). [NLCD 2019: USGS National Land Cover Database, 2019 release](https://developers.google.com/earth-engine/datasets/catalog/USGS_NLCD_RELEASES_2019_REL_NLCD). Official catalogue, relevant full description/example checked25October2026. ## Appendix A. Additional controls and exploratory scores and classifiers These tables retain planned additional controls and the exploratory baseline-cohort/matching scores and classifiers. The exact class-matched condition has different availability. Full paired comparisons, calibration levels, year/region strata, all supports and all seeds are supplied as CSV files. Preplanned additional controls and calendar anchors: | Representation | Task | Score / classifier | True positives / positives | False positives / negatives | J | | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Initial-class controls | 5/21 | 12/54 | 0.016 | | AlphaEarth V1 mosaic | gain | Frozen 2017 reference | 3/24 | 11/85 | -0.004 | | AlphaEarth V1 mosaic | gain | Frozen 2017 + tree controls | 4/24 | 15/85 | -0.010 | | AlphaEarth V1 mosaic | gain | Nearest non-tree controls | 7/24 | 18/85 | 0.080 | | AlphaEarth V1 mosaic | gain | Non-tree controls | 6/24 | 14/85 | 0.085 | | AlphaEarth V1 mosaic | loss | Frozen 2017 reference | 57/124 | 15/213 | 0.389 | | AlphaEarth V1 mosaic | loss | Frozen 2017 + tree controls | 57/124 | 21/213 | 0.361 | | TESSERA v1 | gain | Initial-class controls | 4/21 | 3/54 | 0.135 | | TESSERA v1 | gain | Frozen 2017 reference | 8/24 | 7/85 | 0.251 | | TESSERA v1 | gain | Frozen 2017 + tree controls | 6/24 | 3/85 | 0.215 | | TESSERA v1 | gain | Nearest non-tree controls | 7/24 | 7/85 | 0.209 | | TESSERA v1 | gain | Non-tree controls | 6/24 | 6/85 | 0.179 | | TESSERA v1 | loss | Frozen 2017 reference | 83/124 | 23/213 | 0.561 | | TESSERA v1 | loss | Frozen 2017 + tree controls | 75/124 | 16/213 | 0.530 | Exploratory baseline-cohort and embedding-matched control scores and classifiers: | Representation | Task | Score / classifier | True positives / positives | False positives / negatives | J | | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Baseline-cohort logistic | 5/24 | 5/85 | 0.150 | | AlphaEarth V1 mosaic | gain | Baseline-cohort forest | 7/24 | 8/85 | 0.198 | | AlphaEarth V1 mosaic | gain | Embedding-matched controls | 10/24 | 17/85 | 0.217 | | AlphaEarth V1 mosaic | loss | Baseline-cohort logistic | 35/124 | 21/213 | 0.184 | | AlphaEarth V1 mosaic | loss | Baseline-cohort forest | 83/124 | 38/213 | 0.491 | | AlphaEarth V1 mosaic | loss | Embedding-matched controls | 65/124 | 33/213 | 0.369 | | TESSERA v1 | gain | Baseline-cohort logistic | 18/24 | 16/85 | 0.562 | | TESSERA v1 | gain | Baseline-cohort forest | 15/24 | 4/85 | 0.578 | | TESSERA v1 | gain | Embedding-matched controls | 7/24 | 7/85 | 0.209 | | TESSERA v1 | loss | Baseline-cohort logistic | 79/124 | 22/213 | 0.534 | | TESSERA v1 | loss | Baseline-cohort forest | 90/124 | 34/213 | 0.566 | | TESSERA v1 | loss | Embedding-matched controls | 101/124 | 34/213 | 0.655 | Exploratory baseline-reference access including the target region, with target labels still held out: | Representation | Reference selection | Task | Score / classifier versus fixed start-year reference | J difference | Conditional 95% interval | Changed flags / cases | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | global | gain | Concurrent reference | -0.024 | -0.061 to 0.000 | 2/109 | | AlphaEarth V1 mosaic | global | gain | Control-mean adjustment | 0.001 | -0.093 to 0.117 | 9/109 | | AlphaEarth V1 mosaic | global | loss | Concurrent reference | -0.011 | -0.062 to 0.026 | 11/337 | | AlphaEarth V1 mosaic | global | loss | Control-mean adjustment | 0.002 | -0.037 to 0.038 | 9/337 | | AlphaEarth V1 mosaic | nearest8 | gain | Concurrent reference | -0.041 | -0.193 to 0.128 | 10/109 | | AlphaEarth V1 mosaic | nearest8 | gain | Control-mean adjustment | 0.006 | -0.067 to 0.128 | 4/109 | | AlphaEarth V1 mosaic | nearest8 | loss | Concurrent reference | -0.089 | -0.166 to -0.005 | 47/337 | | AlphaEarth V1 mosaic | nearest8 | loss | Control-mean adjustment | 0.003 | -0.043 to 0.054 | 20/337 | | TESSERA v1 | global | gain | Concurrent reference | -0.012 | -0.063 to 0.040 | 5/109 | | TESSERA v1 | global | gain | Control-mean adjustment | -0.013 | -0.151 to 0.084 | 8/109 | | TESSERA v1 | global | loss | Concurrent reference | -0.046 | -0.101 to 0.009 | 23/337 | | TESSERA v1 | global | loss | Control-mean adjustment | 0.031 | -0.011 to 0.071 | 12/337 | | TESSERA v1 | nearest8 | gain | Concurrent reference | -0.024 | -0.056 to 0.000 | 2/109 | | TESSERA v1 | nearest8 | gain | Control-mean adjustment | 0.000 | 0.000 to 0.000 | 0/109 | | TESSERA v1 | nearest8 | loss | Concurrent reference | -0.033 | -0.071 to 0.000 | 19/337 | | TESSERA v1 | nearest8 | loss | Control-mean adjustment | 0.045 | -0.014 to 0.113 | 24/337 | ![Operating points across exploratory calibration levels](figures/diagnostics/calibration-sensitivity.png) ## Appendix B. The separate NLCD Forest endpoint All scores and classifiers retain LCMAP fitting/calibration. The complete method set, strata, paired contrasts and conditional intervals are in analysis/nlcd-v4, analysis/nlcd-initial-v17, analysis/nlcd-geography-v18 and analysis/nlcd-class-v19. The rows below are fixed primary methods, the standard magnitude/direct/state comparators, and the added input-period diagnostic. No winning condition is selected. Initially supplied classes and repeated plot-years remain part of the design. Magnitude, direct and subsequent-state comparisons under primary labels: | Representation | Pair end years | Task | Score / classifier | True positives / positives | False positives / negatives | J | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 2022–2023 | gain | Cosine change magnitude | 0/7 | 1/115 | -0.009 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Direct gradient boosting | 1/7 | 10/115 | 0.056 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Subsequent-state forest | 0/7 | 1/115 | -0.009 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Cosine change magnitude | 24/26 | 23/100 | 0.693 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Direct gradient boosting | 24/26 | 28/100 | 0.643 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Subsequent-state forest | 21/26 | 37/100 | 0.438 | | TESSERA v1 | 2022–2023 | gain | Cosine change magnitude | 1/7 | 7/115 | 0.082 | | TESSERA v1 | 2022–2023 | gain | Direct gradient boosting | 4/7 | 8/115 | 0.502 | | TESSERA v1 | 2022–2023 | gain | Subsequent-state forest | 3/7 | 2/115 | 0.411 | | TESSERA v1 | 2022–2023 | loss | Cosine change magnitude | 25/26 | 30/100 | 0.662 | | TESSERA v1 | 2022–2023 | loss | Direct gradient boosting | 19/26 | 25/100 | 0.481 | | TESSERA v1 | 2022–2023 | loss | Subsequent-state forest | 24/26 | 34/100 | 0.583 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Cosine change magnitude | 0/24 | 1/121 | -0.008 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Direct gradient boosting | 15/24 | 6/121 | 0.575 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Subsequent-state forest | 8/24 | 3/121 | 0.309 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Cosine change magnitude | 25/29 | 8/100 | 0.782 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Direct gradient boosting | 23/29 | 22/100 | 0.573 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Subsequent-state forest | 25/29 | 37/100 | 0.492 | | TESSERA v1 | 2020–2021 | gain | Cosine change magnitude | 0/24 | 6/121 | -0.050 | | TESSERA v1 | 2020–2021 | gain | Direct gradient boosting | 19/24 | 9/121 | 0.717 | | TESSERA v1 | 2020–2021 | gain | Subsequent-state forest | 16/24 | 8/121 | 0.601 | | TESSERA v1 | 2020–2021 | loss | Cosine change magnitude | 22/29 | 18/100 | 0.579 | | TESSERA v1 | 2020–2021 | loss | Direct gradient boosting | 23/29 | 19/100 | 0.603 | | TESSERA v1 | 2020–2021 | loss | Subsequent-state forest | 28/29 | 26/100 | 0.706 | Initial versus subsequent state forest scores and classifiers under primary labels: | Representation | Pair end years | Task | State score or classifier | True positives / positives | False positives / negatives | J | AUROC | | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 2022–2023 | gain | Baseline-cohort forest | 0/7 | 1/115 | -0.009 | 0.973 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Initial-vector cohort forest | 2/7 | 2/115 | 0.268 | 0.961 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Initial-vector state forest | 1/7 | 1/115 | 0.134 | 0.793 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Subsequent-state forest | 0/7 | 1/115 | -0.009 | 0.843 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Baseline-cohort forest | 18/26 | 29/100 | 0.402 | 0.807 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Initial-vector cohort forest | 3/26 | 35/100 | -0.235 | 0.471 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Initial-vector state forest | 6/26 | 42/100 | -0.189 | 0.423 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Subsequent-state forest | 21/26 | 37/100 | 0.438 | 0.875 | | TESSERA v1 | 2022–2023 | gain | Baseline-cohort forest | 1/7 | 3/115 | 0.117 | 0.850 | | TESSERA v1 | 2022–2023 | gain | Initial-vector cohort forest | 2/7 | 1/115 | 0.277 | 0.866 | | TESSERA v1 | 2022–2023 | gain | Initial-vector state forest | 2/7 | 1/115 | 0.277 | 0.768 | | TESSERA v1 | 2022–2023 | gain | Subsequent-state forest | 3/7 | 2/115 | 0.411 | 0.928 | | TESSERA v1 | 2022–2023 | loss | Baseline-cohort forest | 19/26 | 28/100 | 0.451 | 0.892 | | TESSERA v1 | 2022–2023 | loss | Initial-vector cohort forest | 7/26 | 25/100 | 0.019 | 0.488 | | TESSERA v1 | 2022–2023 | loss | Initial-vector state forest | 11/26 | 38/100 | 0.043 | 0.489 | | TESSERA v1 | 2022–2023 | loss | Subsequent-state forest | 24/26 | 34/100 | 0.583 | 0.918 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Baseline-cohort forest | 6/24 | 6/121 | 0.200 | 0.838 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Initial-vector cohort forest | 6/24 | 6/121 | 0.200 | 0.829 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Initial-vector state forest | 8/24 | 6/121 | 0.284 | 0.750 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Subsequent-state forest | 8/24 | 3/121 | 0.309 | 0.846 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Baseline-cohort forest | 20/29 | 28/100 | 0.410 | 0.723 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Initial-vector cohort forest | 16/29 | 36/100 | 0.192 | 0.594 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Initial-vector state forest | 19/29 | 37/100 | 0.285 | 0.572 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Subsequent-state forest | 25/29 | 37/100 | 0.492 | 0.840 | | TESSERA v1 | 2020–2021 | gain | Baseline-cohort forest | 15/24 | 10/121 | 0.542 | 0.914 | | TESSERA v1 | 2020–2021 | gain | Initial-vector cohort forest | 13/24 | 7/121 | 0.484 | 0.855 | | TESSERA v1 | 2020–2021 | gain | Initial-vector state forest | 16/24 | 10/121 | 0.584 | 0.830 | | TESSERA v1 | 2020–2021 | gain | Subsequent-state forest | 16/24 | 8/121 | 0.601 | 0.898 | | TESSERA v1 | 2020–2021 | loss | Baseline-cohort forest | 25/29 | 26/100 | 0.602 | 0.821 | | TESSERA v1 | 2020–2021 | loss | Initial-vector cohort forest | 9/29 | 25/100 | 0.060 | 0.599 | | TESSERA v1 | 2020–2021 | loss | Initial-vector state forest | 17/29 | 30/100 | 0.286 | 0.592 | | TESSERA v1 | 2020–2021 | loss | Subsequent-state forest | 28/29 | 26/100 | 0.706 | 0.892 | Primary/alternate-agreeing cover subset, reported separately: | Representation | Pair end years | Task | Score / classifier | True positives / positives | False positives / negatives | J | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 2022–2023 | gain | Start-year reference | 0/5 | 11/59 | -0.186 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Concurrent reference | 0/5 | 11/59 | -0.186 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Control-mean adjustment | 0/5 | 13/59 | -0.220 | | AlphaEarth V1 mosaic | 2022–2023 | gain | Cosine change magnitude | 0/5 | 0/59 | 0.000 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Start-year reference | 12/22 | 4/83 | 0.497 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Concurrent reference | 11/22 | 2/83 | 0.476 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Control-mean adjustment | 13/22 | 8/83 | 0.495 | | AlphaEarth V1 mosaic | 2022–2023 | loss | Cosine change magnitude | 20/22 | 17/83 | 0.704 | | TESSERA v1 | 2022–2023 | gain | Start-year reference | 1/5 | 16/59 | -0.071 | | TESSERA v1 | 2022–2023 | gain | Concurrent reference | 1/5 | 19/59 | -0.122 | | TESSERA v1 | 2022–2023 | gain | Control-mean adjustment | 0/5 | 13/59 | -0.220 | | TESSERA v1 | 2022–2023 | gain | Cosine change magnitude | 0/5 | 1/59 | -0.017 | | TESSERA v1 | 2022–2023 | loss | Start-year reference | 21/22 | 20/83 | 0.714 | | TESSERA v1 | 2022–2023 | loss | Concurrent reference | 18/22 | 19/83 | 0.589 | | TESSERA v1 | 2022–2023 | loss | Control-mean adjustment | 18/22 | 9/83 | 0.710 | | TESSERA v1 | 2022–2023 | loss | Cosine change magnitude | 21/22 | 26/83 | 0.641 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Start-year reference | 2/13 | 5/60 | 0.071 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Concurrent reference | 0/13 | 4/60 | -0.067 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Control-mean adjustment | 2/13 | 9/60 | 0.004 | | AlphaEarth V1 mosaic | 2020–2021 | gain | Cosine change magnitude | 0/13 | 1/60 | -0.017 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Start-year reference | 8/23 | 6/85 | 0.277 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Concurrent reference | 7/23 | 7/85 | 0.222 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Control-mean adjustment | 8/23 | 7/85 | 0.265 | | AlphaEarth V1 mosaic | 2020–2021 | loss | Cosine change magnitude | 20/23 | 5/85 | 0.811 | | TESSERA v1 | 2020–2021 | gain | Start-year reference | 2/13 | 14/60 | -0.079 | | TESSERA v1 | 2020–2021 | gain | Concurrent reference | 1/13 | 12/60 | -0.123 | | TESSERA v1 | 2020–2021 | gain | Control-mean adjustment | 0/13 | 5/60 | -0.083 | | TESSERA v1 | 2020–2021 | gain | Cosine change magnitude | 0/13 | 4/60 | -0.067 | | TESSERA v1 | 2020–2021 | loss | Start-year reference | 12/23 | 14/85 | 0.357 | | TESSERA v1 | 2020–2021 | loss | Concurrent reference | 12/23 | 18/85 | 0.310 | | TESSERA v1 | 2020–2021 | loss | Control-mean adjustment | 11/23 | 14/85 | 0.314 | | TESSERA v1 | 2020–2021 | loss | Cosine change magnitude | 18/23 | 16/85 | 0.594 | ## Appendix C. Saved single-date native input sensitivity Only one date per pair is substituted; the other remains a mosaic vector. The fixed-pair native experiment remains incomplete. Identical-score tolerance is an absolute tolerance of 1e-12. Counts below refer to the same small set of pairs across multiple methods and are not independent observations. Full scores, common-case TP/FP counts and input identities are supplied in analysis/native-single-date-v22. | Support | Task | Score / classifier | Cases | Identical substituted inputs | Identical scores | Changed flags | | --- | --- | --- | --- | --- | --- | --- | | centre | gain | Concurrent reference | 2 | 2 | 2 | 0 | | centre | gain | Control-mean adjustment | 2 | 2 | 2 | 0 | | centre | gain | Cosine change magnitude | 2 | 2 | 2 | 0 | | centre | gain | Direct gradient boosting | 2 | 2 | 2 | 0 | | centre | gain | End-year reference | 2 | 2 | 2 | 0 | | centre | gain | Pooled reference | 2 | 2 | 2 | 0 | | centre | gain | Standardized concurrent | 2 | 2 | 2 | 0 | | centre | gain | Start-year reference | 2 | 2 | 2 | 0 | | centre | gain | State random forest | 2 | 2 | 2 | 0 | | centre | gain | Subsequent-reference similarity | 2 | 2 | 2 | 0 | | centre | gain | Subsequent-state logistic | 2 | 2 | 2 | 0 | | centre | gain | Subsequent-state forest | 2 | 2 | 2 | 0 | | centre | loss | Concurrent reference | 18 | 8 | 8 | 2 | | centre | loss | Control-mean adjustment | 18 | 8 | 8 | 1 | | centre | loss | Cosine change magnitude | 18 | 8 | 8 | 0 | | centre | loss | Direct gradient boosting | 18 | 8 | 8 | 1 | | centre | loss | End-year reference | 18 | 8 | 8 | 2 | | centre | loss | Pooled reference | 18 | 8 | 8 | 2 | | centre | loss | Standardized concurrent | 18 | 8 | 8 | 1 | | centre | loss | Start-year reference | 18 | 8 | 8 | 2 | | centre | loss | State random forest | 18 | 8 | 8 | 3 | | centre | loss | Subsequent-reference similarity | 18 | 8 | 11 | 1 | | centre | loss | Subsequent-state logistic | 18 | 8 | 11 | 0 | | centre | loss | Subsequent-state forest | 18 | 8 | 11 | 0 | | patch | gain | Concurrent reference | 2 | 1 | 1 | 0 | | patch | gain | Control-mean adjustment | 2 | 1 | 1 | 0 | | patch | gain | Cosine change magnitude | 2 | 1 | 1 | 0 | | patch | gain | Direct gradient boosting | 2 | 1 | 1 | 0 | | patch | gain | End-year reference | 2 | 1 | 1 | 0 | | patch | gain | Pooled reference | 2 | 1 | 1 | 0 | | patch | gain | Standardized concurrent | 2 | 1 | 1 | 0 | | patch | gain | Start-year reference | 2 | 1 | 1 | 0 | | patch | gain | State random forest | 2 | 1 | 1 | 0 | | patch | gain | Subsequent-reference similarity | 2 | 1 | 1 | 0 | | patch | gain | Subsequent-state logistic | 2 | 1 | 1 | 0 | | patch | gain | Subsequent-state forest | 2 | 1 | 1 | 0 | | patch | loss | Concurrent reference | 18 | 6 | 6 | 1 | | patch | loss | Control-mean adjustment | 18 | 6 | 6 | 1 | | patch | loss | Cosine change magnitude | 18 | 6 | 6 | 0 | | patch | loss | Direct gradient boosting | 18 | 6 | 9 | 0 | | patch | loss | End-year reference | 18 | 6 | 6 | 1 | | patch | loss | Pooled reference | 18 | 6 | 6 | 0 | | patch | loss | Standardized concurrent | 18 | 6 | 6 | 2 | | patch | loss | Start-year reference | 18 | 6 | 6 | 2 | | patch | loss | State random forest | 18 | 6 | 6 | 0 | | patch | loss | Subsequent-reference similarity | 18 | 6 | 10 | 1 | | patch | loss | Subsequent-state logistic | 18 | 6 | 10 | 0 | | patch | loss | Subsequent-state forest | 18 | 6 | 10 | 0 | ## Appendix D. Matched direct-learner input comparisons ### Retained paired contrasts: Matched direct-learner input comparison: exploratory The initial-vector state diagnostic does not hold the training target and algorithm constant against the direct transition comparator. The later exploratory matched-learner analysis addresses that comparison: before-only, after-only, difference-only and original both-date inputs use the same earlier transition labels, learner hyperparameters, partitions and calibration rule. Input dimensions differ, so equal settings do not imply equal capacity. Original both-date scores/flags/thresholds are reconstructed exactly. Both endpoints had already been inspected before this amendment; the Forest results here are reused exploratory data, not a fresh validation. The before-only input retains gain discrimination in the observed sample. Difference-only J is lower than both-date J in 6/6 gain conditions and higher in 6/6 loss conditions across the three separately reported populations and two products. These are descriptive point directions on shared fitted samples; conditional intervals often include zero, and no universal ordering or equivalence is established. A learned difference-vector score or classifier can flag some LCMAP gains even when scalar magnitude flags none. Conversely, differencing alone does not give uniformly strong gain transfer. These results qualify interpretations based solely on either a magnitude score or a subsequent-state classifier. ![Matched direct-learner input comparisons](figures/direct-inputs.png) Appendix D retains every input condition's operating point. The original LCMAP reduced-input contrasts below show uncertainty and changed decisions explicitly; the separate Forest contrasts and label-agreement subsets are in the supplied CSV files. Before-only results remain retrospective associations with dated vector values, not prospective forecasts or identification of the source of the information. | Representation | Task | Input versus both-date direct learner | J difference | Conditional 95% interval | Changed flags / cases | | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Direct learner: before only | 0.006 | -0.240 to 0.217 | 14/109 | | AlphaEarth V1 mosaic | loss | Direct learner: before only | -0.519 | -0.687 to -0.355 | 123/337 | | TESSERA v1 | gain | Direct learner: before only | -0.082 | -0.229 to 0.080 | 15/109 | | TESSERA v1 | loss | Direct learner: before only | -0.486 | -0.585 to -0.379 | 102/337 | | AlphaEarth V1 mosaic | gain | Direct learner: after only | 0.043 | -0.348 to 0.346 | 22/109 | | AlphaEarth V1 mosaic | loss | Direct learner: after only | -0.289 | -0.384 to -0.193 | 71/337 | | TESSERA v1 | gain | Direct learner: after only | -0.048 | -0.155 to 0.057 | 9/109 | | TESSERA v1 | loss | Direct learner: after only | -0.051 | -0.124 to 0.024 | 44/337 | | AlphaEarth V1 mosaic | gain | Direct learner: difference only | -0.119 | -0.341 to 0.115 | 17/109 | | AlphaEarth V1 mosaic | loss | Direct learner: difference only | 0.040 | -0.053 to 0.139 | 51/337 | | TESSERA v1 | gain | Direct learner: difference only | -0.313 | -0.609 to 0.043 | 41/109 | | TESSERA v1 | loss | Direct learner: difference only | 0.007 | -0.080 to 0.106 | 45/337 | All four fixed inputs are shown. Both-date input means the original concatenation of before, after and difference vectors. This amendment was fixed after inspecting both endpoints; all rows are exploratory and share earlier LCMAP fitting/calibration. Full conditional paired intervals, year/region strata, feature dimensions and threshold/reconstruction checks are in analysis/direct-inputs-v23 and analysis/nlcd-direct-inputs-v23. No winning input condition is selected. | Test population | Representation | Task | Direct input condition | True positives / positives | False positives / negatives | J | AUROC | | --- | --- | --- | --- | --- | --- | --- | --- | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct learner: after only | 15/24 | 13/85 | 0.472 | 0.861 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct learner: before only | 13/24 | 9/85 | 0.436 | 0.812 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct learner: difference only | 10/24 | 9/85 | 0.311 | 0.729 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct gradient boosting | 12/24 | 6/85 | 0.429 | 0.866 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct learner: after only | 74/124 | 31/213 | 0.451 | 0.847 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct learner: before only | 42/124 | 25/213 | 0.221 | 0.702 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct learner: difference only | 113/124 | 28/213 | 0.780 | 0.953 | | LCMAP, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct gradient boosting | 111/124 | 33/213 | 0.740 | 0.957 | | LCMAP, end 2020–2021 | TESSERA v1 | gain | Direct learner: after only | 12/24 | 5/85 | 0.441 | 0.876 | | LCMAP, end 2020–2021 | TESSERA v1 | gain | Direct learner: before only | 14/24 | 15/85 | 0.407 | 0.853 | | LCMAP, end 2020–2021 | TESSERA v1 | gain | Direct learner: difference only | 11/24 | 24/85 | 0.176 | 0.621 | | LCMAP, end 2020–2021 | TESSERA v1 | gain | Direct gradient boosting | 14/24 | 8/85 | 0.489 | 0.874 | | LCMAP, end 2020–2021 | TESSERA v1 | loss | Direct learner: after only | 110/124 | 34/213 | 0.727 | 0.932 | | LCMAP, end 2020–2021 | TESSERA v1 | loss | Direct learner: before only | 59/124 | 39/213 | 0.293 | 0.764 | | LCMAP, end 2020–2021 | TESSERA v1 | loss | Direct learner: difference only | 109/124 | 20/213 | 0.785 | 0.948 | | LCMAP, end 2020–2021 | TESSERA v1 | loss | Direct gradient boosting | 114/124 | 30/213 | 0.779 | 0.959 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | gain | Direct learner: after only | 3/7 | 7/115 | 0.368 | 0.834 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | gain | Direct learner: before only | 3/7 | 6/115 | 0.376 | 0.831 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | gain | Direct learner: difference only | 1/7 | 18/115 | -0.014 | 0.544 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | gain | Direct gradient boosting | 1/7 | 10/115 | 0.056 | 0.793 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | loss | Direct learner: after only | 19/26 | 27/100 | 0.461 | 0.844 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | loss | Direct learner: before only | 8/26 | 14/100 | 0.168 | 0.670 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | loss | Direct learner: difference only | 25/26 | 26/100 | 0.702 | 0.940 | | NLCD Forest, end 2022–2023 | AlphaEarth V1 mosaic | loss | Direct gradient boosting | 24/26 | 28/100 | 0.643 | 0.920 | | NLCD Forest, end 2022–2023 | TESSERA v1 | gain | Direct learner: after only | 3/7 | 3/115 | 0.402 | 0.771 | | NLCD Forest, end 2022–2023 | TESSERA v1 | gain | Direct learner: before only | 4/7 | 14/115 | 0.450 | 0.831 | | NLCD Forest, end 2022–2023 | TESSERA v1 | gain | Direct learner: difference only | 1/7 | 35/115 | -0.161 | 0.328 | | NLCD Forest, end 2022–2023 | TESSERA v1 | gain | Direct gradient boosting | 4/7 | 8/115 | 0.502 | 0.860 | | NLCD Forest, end 2022–2023 | TESSERA v1 | loss | Direct learner: after only | 22/26 | 29/100 | 0.556 | 0.867 | | NLCD Forest, end 2022–2023 | TESSERA v1 | loss | Direct learner: before only | 7/26 | 17/100 | 0.099 | 0.535 | | NLCD Forest, end 2022–2023 | TESSERA v1 | loss | Direct learner: difference only | 22/26 | 19/100 | 0.656 | 0.882 | | NLCD Forest, end 2022–2023 | TESSERA v1 | loss | Direct gradient boosting | 19/26 | 25/100 | 0.481 | 0.868 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct learner: after only | 13/24 | 12/121 | 0.442 | 0.890 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct learner: before only | 10/24 | 8/121 | 0.351 | 0.857 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct learner: difference only | 8/24 | 13/121 | 0.226 | 0.722 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | gain | Direct gradient boosting | 15/24 | 6/121 | 0.575 | 0.917 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct learner: after only | 19/29 | 24/100 | 0.415 | 0.790 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct learner: before only | 6/29 | 19/100 | 0.017 | 0.609 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct learner: difference only | 24/29 | 13/100 | 0.698 | 0.903 | | NLCD Forest, end 2020–2021 | AlphaEarth V1 mosaic | loss | Direct gradient boosting | 23/29 | 22/100 | 0.573 | 0.884 | | NLCD Forest, end 2020–2021 | TESSERA v1 | gain | Direct learner: after only | 11/24 | 6/121 | 0.409 | 0.914 | | NLCD Forest, end 2020–2021 | TESSERA v1 | gain | Direct learner: before only | 17/24 | 17/121 | 0.568 | 0.903 | | NLCD Forest, end 2020–2021 | TESSERA v1 | gain | Direct learner: difference only | 8/24 | 48/121 | -0.063 | 0.446 | | NLCD Forest, end 2020–2021 | TESSERA v1 | gain | Direct gradient boosting | 19/24 | 9/121 | 0.717 | 0.936 | | NLCD Forest, end 2020–2021 | TESSERA v1 | loss | Direct learner: after only | 23/29 | 32/100 | 0.473 | 0.804 | | NLCD Forest, end 2020–2021 | TESSERA v1 | loss | Direct learner: before only | 16/29 | 27/100 | 0.282 | 0.621 | | NLCD Forest, end 2020–2021 | TESSERA v1 | loss | Direct learner: difference only | 23/29 | 9/100 | 0.703 | 0.913 | | NLCD Forest, end 2020–2021 | TESSERA v1 | loss | Direct gradient boosting | 23/29 | 19/100 | 0.603 | 0.861 | ## Appendix E. Reference/control availability and longer reference intervals Post-review descriptive diagnostics. All counts refer to baseline-selected plots outside the named fold, not verified later stability. Full one-year/individual-plot rows and same-membership means are retained in analysis/review-diagnostics-r1. No multi-year target screening is inferred. | Representation | Held-out region | Year | Assigned references outside region | Valid references | Assigned controls outside region | Valid controls | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 0 | 2017 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 0 | 2018 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 0 | 2019 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 0 | 2020 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 0 | 2021 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 0 | 2022 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 0 | 2023 | 53 | 53 | 51 | 51 | | AlphaEarth V1 mosaic | 1 | 2017 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 1 | 2018 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 1 | 2019 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 1 | 2020 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 1 | 2021 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 1 | 2022 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 1 | 2023 | 49 | 49 | 46 | 46 | | AlphaEarth V1 mosaic | 2 | 2017 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 2 | 2018 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 2 | 2019 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 2 | 2020 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 2 | 2021 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 2 | 2022 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 2 | 2023 | 48 | 48 | 53 | 53 | | AlphaEarth V1 mosaic | 3 | 2017 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 3 | 2018 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 3 | 2019 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 3 | 2020 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 3 | 2021 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 3 | 2022 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 3 | 2023 | 39 | 39 | 38 | 38 | | AlphaEarth V1 mosaic | 4 | 2017 | 43 | 43 | 44 | 44 | | AlphaEarth V1 mosaic | 4 | 2018 | 43 | 43 | 44 | 44 | | AlphaEarth V1 mosaic | 4 | 2019 | 43 | 43 | 44 | 44 | | AlphaEarth V1 mosaic | 4 | 2020 | 43 | 43 | 44 | 44 | | AlphaEarth V1 mosaic | 4 | 2021 | 43 | 43 | 44 | 44 | | AlphaEarth V1 mosaic | 4 | 2022 | 43 | 43 | 44 | 44 | | AlphaEarth V1 mosaic | 4 | 2023 | 43 | 43 | 44 | 44 | | TESSERA v1 | 0 | 2017 | 53 | 53 | 51 | 50 | | TESSERA v1 | 0 | 2018 | 53 | 51 | 51 | 47 | | TESSERA v1 | 0 | 2019 | 53 | 53 | 51 | 50 | | TESSERA v1 | 0 | 2020 | 53 | 53 | 51 | 50 | | TESSERA v1 | 0 | 2021 | 53 | 53 | 51 | 50 | | TESSERA v1 | 0 | 2022 | 53 | 53 | 51 | 50 | | TESSERA v1 | 0 | 2023 | 53 | 53 | 51 | 50 | | TESSERA v1 | 1 | 2017 | 49 | 49 | 46 | 45 | | TESSERA v1 | 1 | 2018 | 49 | 47 | 46 | 42 | | TESSERA v1 | 1 | 2019 | 49 | 49 | 46 | 45 | | TESSERA v1 | 1 | 2020 | 49 | 49 | 46 | 45 | | TESSERA v1 | 1 | 2021 | 49 | 49 | 46 | 45 | | TESSERA v1 | 1 | 2022 | 49 | 49 | 46 | 45 | | TESSERA v1 | 1 | 2023 | 49 | 49 | 46 | 45 | | TESSERA v1 | 2 | 2017 | 48 | 48 | 53 | 52 | | TESSERA v1 | 2 | 2018 | 48 | 46 | 53 | 49 | | TESSERA v1 | 2 | 2019 | 48 | 48 | 53 | 52 | | TESSERA v1 | 2 | 2020 | 48 | 48 | 53 | 52 | | TESSERA v1 | 2 | 2021 | 48 | 48 | 53 | 52 | | TESSERA v1 | 2 | 2022 | 48 | 48 | 53 | 52 | | TESSERA v1 | 2 | 2023 | 48 | 48 | 53 | 52 | | TESSERA v1 | 3 | 2017 | 39 | 39 | 38 | 37 | | TESSERA v1 | 3 | 2018 | 39 | 37 | 38 | 34 | | TESSERA v1 | 3 | 2019 | 39 | 39 | 38 | 37 | | TESSERA v1 | 3 | 2020 | 39 | 39 | 38 | 37 | | TESSERA v1 | 3 | 2021 | 39 | 39 | 38 | 37 | | TESSERA v1 | 3 | 2022 | 39 | 39 | 38 | 37 | | TESSERA v1 | 3 | 2023 | 39 | 39 | 38 | 37 | | TESSERA v1 | 4 | 2017 | 43 | 43 | 44 | 44 | | TESSERA v1 | 4 | 2018 | 43 | 43 | 44 | 44 | | TESSERA v1 | 4 | 2019 | 43 | 43 | 44 | 44 | | TESSERA v1 | 4 | 2020 | 43 | 43 | 44 | 44 | | TESSERA v1 | 4 | 2021 | 43 | 43 | 44 | 44 | | TESSERA v1 | 4 | 2022 | 43 | 43 | 44 | 44 | | TESSERA v1 | 4 | 2023 | 43 | 43 | 44 | 44 | Descriptive 2017-to-later reference intervals: | Representation | Held-out region | Reference interval | References before | References after | Paired controls | Centroid dissimilarity | Median control dissimilarity | | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | 0 | 2017–2019 | 53 | 53 | 51 | 0.0098 | 0.0297 | | AlphaEarth V1 mosaic | 0 | 2017–2020 | 53 | 53 | 51 | 0.0065 | 0.0239 | | AlphaEarth V1 mosaic | 0 | 2017–2021 | 53 | 53 | 51 | 0.0068 | 0.0297 | | AlphaEarth V1 mosaic | 0 | 2017–2022 | 53 | 53 | 51 | 0.0112 | 0.0366 | | AlphaEarth V1 mosaic | 0 | 2017–2023 | 53 | 53 | 51 | 0.0101 | 0.0339 | | AlphaEarth V1 mosaic | 1 | 2017–2019 | 49 | 49 | 46 | 0.0104 | 0.0258 | | AlphaEarth V1 mosaic | 1 | 2017–2020 | 49 | 49 | 46 | 0.0067 | 0.0263 | | AlphaEarth V1 mosaic | 1 | 2017–2021 | 49 | 49 | 46 | 0.0078 | 0.0275 | | AlphaEarth V1 mosaic | 1 | 2017–2022 | 49 | 49 | 46 | 0.0133 | 0.0316 | | AlphaEarth V1 mosaic | 1 | 2017–2023 | 49 | 49 | 46 | 0.0105 | 0.0330 | | AlphaEarth V1 mosaic | 2 | 2017–2019 | 48 | 48 | 53 | 0.0089 | 0.0253 | | AlphaEarth V1 mosaic | 2 | 2017–2020 | 48 | 48 | 53 | 0.0056 | 0.0255 | | AlphaEarth V1 mosaic | 2 | 2017–2021 | 48 | 48 | 53 | 0.0069 | 0.0282 | | AlphaEarth V1 mosaic | 2 | 2017–2022 | 48 | 48 | 53 | 0.0116 | 0.0334 | | AlphaEarth V1 mosaic | 2 | 2017–2023 | 48 | 48 | 53 | 0.0095 | 0.0333 | | AlphaEarth V1 mosaic | 3 | 2017–2019 | 39 | 39 | 38 | 0.0118 | 0.0247 | | AlphaEarth V1 mosaic | 3 | 2017–2020 | 39 | 39 | 38 | 0.0056 | 0.0266 | | AlphaEarth V1 mosaic | 3 | 2017–2021 | 39 | 39 | 38 | 0.0087 | 0.0283 | | AlphaEarth V1 mosaic | 3 | 2017–2022 | 39 | 39 | 38 | 0.0127 | 0.0366 | | AlphaEarth V1 mosaic | 3 | 2017–2023 | 39 | 39 | 38 | 0.0092 | 0.0333 | | AlphaEarth V1 mosaic | 4 | 2017–2019 | 43 | 43 | 44 | 0.0160 | 0.0325 | | AlphaEarth V1 mosaic | 4 | 2017–2020 | 43 | 43 | 44 | 0.0086 | 0.0280 | | AlphaEarth V1 mosaic | 4 | 2017–2021 | 43 | 43 | 44 | 0.0079 | 0.0355 | | AlphaEarth V1 mosaic | 4 | 2017–2022 | 43 | 43 | 44 | 0.0177 | 0.0371 | | AlphaEarth V1 mosaic | 4 | 2017–2023 | 43 | 43 | 44 | 0.0153 | 0.0366 | | TESSERA v1 | 0 | 2017–2019 | 53 | 53 | 50 | 0.0541 | 0.1426 | | TESSERA v1 | 0 | 2017–2020 | 53 | 53 | 50 | 0.0601 | 0.1537 | | TESSERA v1 | 0 | 2017–2021 | 53 | 53 | 50 | 0.0720 | 0.1547 | | TESSERA v1 | 0 | 2017–2022 | 53 | 53 | 50 | 0.0835 | 0.1735 | | TESSERA v1 | 0 | 2017–2023 | 53 | 53 | 50 | 0.0758 | 0.1739 | | TESSERA v1 | 1 | 2017–2019 | 49 | 49 | 45 | 0.0566 | 0.1429 | | TESSERA v1 | 1 | 2017–2020 | 49 | 49 | 45 | 0.0626 | 0.1625 | | TESSERA v1 | 1 | 2017–2021 | 49 | 49 | 45 | 0.0748 | 0.1627 | | TESSERA v1 | 1 | 2017–2022 | 49 | 49 | 45 | 0.0845 | 0.1773 | | TESSERA v1 | 1 | 2017–2023 | 49 | 49 | 45 | 0.0790 | 0.1741 | | TESSERA v1 | 2 | 2017–2019 | 48 | 48 | 52 | 0.0543 | 0.1390 | | TESSERA v1 | 2 | 2017–2020 | 48 | 48 | 52 | 0.0609 | 0.1481 | | TESSERA v1 | 2 | 2017–2021 | 48 | 48 | 52 | 0.0697 | 0.1547 | | TESSERA v1 | 2 | 2017–2022 | 48 | 48 | 52 | 0.0840 | 0.1735 | | TESSERA v1 | 2 | 2017–2023 | 48 | 48 | 52 | 0.0722 | 0.1706 | | TESSERA v1 | 3 | 2017–2019 | 39 | 39 | 37 | 0.0523 | 0.1516 | | TESSERA v1 | 3 | 2017–2020 | 39 | 39 | 37 | 0.0625 | 0.1568 | | TESSERA v1 | 3 | 2017–2021 | 39 | 39 | 37 | 0.0774 | 0.1833 | | TESSERA v1 | 3 | 2017–2022 | 39 | 39 | 37 | 0.0808 | 0.1903 | | TESSERA v1 | 3 | 2017–2023 | 39 | 39 | 37 | 0.0790 | 0.1889 | | TESSERA v1 | 4 | 2017–2019 | 43 | 43 | 44 | 0.0530 | 0.1300 | | TESSERA v1 | 4 | 2017–2020 | 43 | 43 | 44 | 0.0560 | 0.1403 | | TESSERA v1 | 4 | 2017–2021 | 43 | 43 | 44 | 0.0654 | 0.1650 | | TESSERA v1 | 4 | 2017–2022 | 43 | 43 | 44 | 0.0831 | 0.1664 | | TESSERA v1 | 4 | 2017–2023 | 43 | 43 | 44 | 0.0721 | 0.1625 | ## Appendix F. Retained exploratory and unfinished checks ### Endpoint composition and reference availability The gain-negative definition includes 21 later Grass/Shrub transitions. Removing them from this separate calibration/test condition leaves 64 unchanged-cover negatives and 24 gain positives. Cosine change magnitude still detects 0 gains for AlphaEarth; the full two-product table is analysis/endpoint-v13/stable-negative-gain-table.md. Fitted scores are unchanged on all common cases. Within this sample, the floor persists after removing those transitions. Its biological, temporal or spatial-support cause remains unresolved. Before/after similarity components reconstructed from saved scores agree with direct raw-vector calculations to a maximum absolute error of 5.55e-16. The scalar plot shows where the change score or classifier differs from an absolute subsequent-state score. It does not measure ecological condition or establish that either the labels or representations are wrong. ![Before and after cosine similarity to the start-year reference](analysis/endpoint-v13/reference-similarity.png) The same-reference-plots check changes the TESSERA gain concurrent-reference J contrast from -0.012 to +0.024. Target test scores remain fixed; earlier calibration inputs and thresholds change. Original separately-valid reference means remain primary. This sensitivity limits claims about a general advantage or disadvantage of concurrent references. ### Additional controls: Alternative controls, calendar anchors and scores and classifiers The preplanned non-tree, initial-class and frozen-calendar controls produce different operating points (Appendix A). Exact initial-class matching has fewer available cases, so its contrast must use the common-case denominators in the paired comparison file. A gain in its J cannot be compared by subtracting the full-sample fixed start-year reference J. No control rule gives a consistent advantage across all products, tasks and available cases. The preplanned reference allocations, centre/patch supports and global/nearest-reference conditions give the ranges below. These settings reuse the same targets and pool; their signs are sensitivity counts, not independent evidence or a vote for a method. All settings remain available in the machine-readable tables. | Representation | Task | Score / classifier versus fixed start-year reference | Sensitivity settings | Minimum J difference | Maximum J difference | Positive | Negative | Zero | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Concurrent reference | 20 | -0.095 | 0.119 | 9 | 10 | 1 | | AlphaEarth V1 mosaic | gain | Control-mean adjustment | 20 | -0.119 | 0.141 | 13 | 7 | 0 | | AlphaEarth V1 mosaic | loss | Concurrent reference | 20 | -0.169 | 0.003 | 1 | 19 | 0 | | AlphaEarth V1 mosaic | loss | Control-mean adjustment | 20 | -0.040 | 0.052 | 6 | 14 | 0 | | TESSERA v1 | gain | Concurrent reference | 20 | -0.284 | 0.059 | 4 | 15 | 1 | | TESSERA v1 | gain | Control-mean adjustment | 20 | -0.167 | 0.148 | 13 | 6 | 1 | | TESSERA v1 | loss | Concurrent reference | 20 | -0.080 | 0.026 | 6 | 14 | 0 | | TESSERA v1 | loss | Control-mean adjustment | 20 | -0.053 | 0.029 | 9 | 10 | 1 | The original strict reference-region exclusion places the eighth-nearest reference a median 1,071 km away in the centre-pixel sample. An exploratory condition allowing baseline reference/control plots from all regions reduces that median to 579 km. Target fitting and calibration remain region-held-out; the reference-label information access changes. The concurrent-reference J contrast is negative in 8/8 centre conditions across global/nearest selection, while control effects vary (Appendix A). Supervised scores and flags are unchanged by construction and verified. References remain geographically broad and are not locally verified reference ecosystems, even under the relaxed access condition. Exploratory baseline-cohort state classifiers use only the baseline reference/tree-control/non-tree-control cohorts for state fitting, while their thresholds still require earlier target-negative labels. They therefore reduce one source of state-fitting labels without establishing equal annotation budgets or zero transition-label dependence. Embedding-matched controls use initial-vector similarity rather than geographic distance. Their complete results are in Appendix A. Calibration at the additional five- and twenty-percent nominal levels changes some primary contrast directions; the original ten-percent rule remains primary. The auxiliary classification release supplies 109 later labels among 166 baseline-cohort plots, with 0 known binary disagreements and 57 unknown later labels. These are isolated annual observations, not complete histories. No primary plot is removed using those later labels. ### Input-period, coordinate and class checks: What requires the subsequent target vector? Exploratory diagnostics Some gain discrimination remains when a state classifier receives only the initial target vector. The table compares the same fitted state model applied to the before and after vectors, each calibrated using its own earlier-negative scores. Both comparisons retain the same availability mask, which requires both products in both years. The state models were fitted using earlier labelled states, and the encoders and labels were produced retrospectively. These results are therefore associations in a dated representation, not prospective forecasts. | Representation | Task | State fit | Before vector TP; FP | After vector TP; FP | Before J | After J | Before minus after J | Conditional 95% interval | Changed flags / cases | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Initial-vector state forest | 11/24; 9/85 | 9/24; 9/85 | 0.352 | 0.269 | 0.083 | -0.093 to 0.221 | 8/109 | | AlphaEarth V1 mosaic | loss | Initial-vector state forest | 47/124; 56/213 | 113/124; 49/213 | 0.116 | 0.681 | -0.565 | -0.693 to -0.386 | 77/337 | | TESSERA v1 | gain | Initial-vector state forest | 14/24; 10/85 | 14/24; 5/85 | 0.466 | 0.525 | -0.059 | -0.200 to 0.070 | 11/109 | | TESSERA v1 | loss | Initial-vector state forest | 38/124; 49/213 | 119/124; 45/213 | 0.076 | 0.748 | -0.672 | -0.795 to -0.492 | 93/337 | | AlphaEarth V1 mosaic | gain | Initial-vector cohort forest | 7/24; 6/85 | 7/24; 8/85 | 0.221 | 0.198 | 0.024 | -0.092 to 0.148 | 4/109 | | AlphaEarth V1 mosaic | loss | Initial-vector cohort forest | 44/124; 52/213 | 83/124; 38/213 | 0.111 | 0.491 | -0.380 | -0.471 to -0.268 | 63/337 | | TESSERA v1 | gain | Initial-vector cohort forest | 11/24; 9/85 | 15/24; 4/85 | 0.352 | 0.578 | -0.225 | -0.441 to 0.066 | 15/109 | | TESSERA v1 | loss | Initial-vector cohort forest | 26/124; 26/213 | 90/124; 34/213 | 0.088 | 0.566 | -0.479 | -0.569 to -0.374 | 78/337 | Changing the subsequent target-vector values in one fixed within-year permutation leaves all 3568 initial-only scored rows unchanged across 446 target keys. This verifies input dependence in the implementation; it is not another independent empirical replication. Initial reference similarity and baseline-cohort logistic counterparts, including their uncertainty, are retained in analysis/initial-vector-v17/initial-versus-subsequent.csv. The after vector improves several gain scores and classifiers and improves the full state forest substantially for losses. Thus these results do not support the claim that all apparent change accuracy comes from initial information. They do show why subsequent-state classification accuracy alone cannot establish how much information about a transition was newly observed after the initial date. The separate coordinate-only control uses four sine/cosine latitude/longitude features, with no embedding values, year or initial vegetation subclass as learner inputs. It yields the following associations on the same conditional sample. This is one geographic baseline, not two product replications. It cannot identify the source of information in the embeddings; weaker results would not exclude richer static landscape information. | Task | Coordinate-only score or classifier | True positives / positives | False positives / negatives | J | AUROC | | --- | --- | --- | --- | --- | --- | | gain | Coordinate transition boosting | 7/24 | 5/85 | 0.233 | 0.746 | | gain | Coordinate state forest | 1/24 | 1/85 | 0.030 | 0.680 | | gain | Coordinate state logistic | 4/24 | 1/85 | 0.155 | 0.722 | | loss | Coordinate transition boosting | 91/124 | 88/213 | 0.321 | 0.656 | | loss | Coordinate state forest | 45/124 | 50/213 | 0.128 | 0.568 | | loss | Coordinate state logistic | 40/124 | 46/213 | 0.107 | 0.693 | The exploratory initial-class-rate baseline flags 0/24 gains and 0/85 negatives; 62 gain-task cases tie their regional threshold. This simple baseline does not explain the initial-vector forest results: gains are flagged within both initial vegetation classes, as shown below. The original mixed-class thresholds are retained within each subgroup. Small subgroup denominators and richer interactions with landscape context remain limitations. | Representation | Initial vegetation class | State score or classifier | True positives / positives | False positives / negatives | J | AUROC | | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | Grass/forb/herb | Initial-vector state forest | 3/9 | 4/40 | 0.233 | 0.794 | | AlphaEarth V1 mosaic | Grass/forb/herb | Subsequent-state forest | 3/9 | 5/40 | 0.208 | 0.861 | | TESSERA v1 | Grass/forb/herb | Initial-vector state forest | 6/9 | 5/40 | 0.542 | 0.883 | | TESSERA v1 | Grass/forb/herb | Subsequent-state forest | 6/9 | 4/40 | 0.567 | 0.931 | | AlphaEarth V1 mosaic | Shrubs | Initial-vector state forest | 8/15 | 5/45 | 0.422 | 0.769 | | AlphaEarth V1 mosaic | Shrubs | Subsequent-state forest | 6/15 | 4/45 | 0.311 | 0.837 | | TESSERA v1 | Shrubs | Initial-vector state forest | 8/15 | 5/45 | 0.422 | 0.744 | | TESSERA v1 | Shrubs | Subsequent-state forest | 8/15 | 1/45 | 0.511 | 0.834 | All 337 loss-task cases begin as Trees. The class-rate score is constant within each regional fit, making zero loss flags a consequence of the task and threshold definition. This is not counted as a discovery. Region-specific estimates can differ, so the pooled score is not globally constant. ### Native-grid diagnostics Of 120 fixed sampled point-years, 113 native patches were obtained, 113 centres are valid in both inputs and 49 raw centre vectors are exactly equal. Overall summaries are below; role/year/latitude strata and every selected slot remain in analysis/native-v5. Pixel-centre distances are in metres. Centre/patch cosine values compare representations at the same requested coordinates but potentially different footprints. | Native/mosaic co-location metric | Valid pairs | Minimum | Fifth percentile | Median | Maximum | | --- | --- | --- | --- | --- | --- | | Centre cosine | 113 | 0.907660 | 0.982823 | 0.999399 | 1.000000 | | Patch cosine | 113 | 0.977003 | 0.991166 | 0.999435 | 1.000000 | | Pixel-centre distance (m) | 113 | 0.304594 | 1.364329 | 5.021877 | 9.867387 | This diagnostic neither re-fits the full study on native inputs nor tests stability of every small threshold contrast. Any disagreement confines the AlphaEarth estimates to the geographic mosaic actually evaluated. Cross-product differences cannot be attributed solely to encoders with their different input-processing paths. The separately fixed paired-native case check is unfinished: only 1/32 point-years are available, leaving 0/16 complete pairs. Its empty counts are not evidence of unchanged decisions. Checking all saved original targets found no complete native pair, so the full paired comparison cannot be recovered from these measurements without additional acquisition. An exploratory saved-data alternative replaces one available target date in each of 20 pairs, leaving its other date as a mosaic input. It contains 1 gain positive/1 gain negative and 10 loss positive/8 loss negatives; 6 substitutions affect the before date and 14 the after date. All fitting, references, controls and calibration remain mosaics. At least one flag changes in 5/20 pairs across the retained scores and classifiers and supports. This is an availability-selected, mixed-footprint case test, not an estimate of full native-grid performance or which footprint is correct. Identity counts added during self-check show 10/20 identical centre vectors and 7/20 identical patch vectors. Both gain-case centre inputs are identical, so their unchanged centre decisions are predictable, not robustness evidence. Subsequent-state scores can also stay fixed when only the initial date changes. Appendix C gives these counts for every score or classifier and separates score identity from changed flags. The observed changes reinforce the geographic-mosaic scope of small AlphaEarth contrasts; paired-native transfer remains untested. ### Partition sensitivity: Earlier fitting/calibration partition sensitivity: exploratory The final exploratory check completed all 10 fixed earlier geographic-cell assignments: the original partition and its 9 perturbations. Reference cohorts, model settings, later tests and the calibration rule remain fixed. The original condition reconstructs 21408 prediction rows and 480 thresholds. Unsupervised score values on common keys are unchanged; their transferred decisions can vary through calibration. Learned scores can also vary through fitting. The table retains every centre primary correction and direct-learner contrast. Its range is over these overlapping conditions, not an uncertainty interval or a selection of a best split. Positive/negative direction counts are not independent replications. Case ranges show whether differing availability could explain contrasts. There are 4800/4800 available method/fold conditions across all scores and classifiers/supports; 0 lack a threshold record and 0 have invalid calibration. Unavailable conditions are retained in the output rather than replaced. | Representation | Task | Score / classifier versus fixed start-year reference | Available partitions | J difference range | Positive/zero/negative directions | Positive / negative case ranges | Changed flags range | | --- | --- | --- | --- | --- | --- | --- | --- | | AlphaEarth V1 mosaic | gain | Concurrent reference | 10 | -0.035 to +0.065 | 5/3/2 | 24–24 / 85–85 | 0–6 | | AlphaEarth V1 mosaic | gain | Control-mean adjustment | 10 | -0.053 to +0.030 | 2/5/3 | 24–24 / 85–85 | 0–9 | | AlphaEarth V1 mosaic | gain | Direct gradient boosting | 10 | +0.310 to +0.643 | 10/0/0 | 24–24 / 85–85 | 14–41 | | AlphaEarth V1 mosaic | loss | Concurrent reference | 10 | -0.055 to +0.015 | 1/0/9 | 124–124 / 213–213 | 16–34 | | AlphaEarth V1 mosaic | loss | Control-mean adjustment | 10 | -0.069 to +0.026 | 1/0/9 | 124–124 / 213–213 | 6–24 | | AlphaEarth V1 mosaic | loss | Direct gradient boosting | 10 | +0.191 to +0.352 | 10/0/0 | 124–124 / 213–213 | 69–134 | | TESSERA v1 | gain | Concurrent reference | 10 | -0.071 to +0.072 | 2/1/7 | 24–24 / 85–85 | 0–10 | | TESSERA v1 | gain | Control-mean adjustment | 10 | -0.072 to +0.065 | 6/1/3 | 24–24 / 85–85 | 1–10 | | TESSERA v1 | gain | Direct gradient boosting | 10 | +0.256 to +0.484 | 10/0/0 | 24–24 / 85–85 | 26–52 | | TESSERA v1 | loss | Concurrent reference | 10 | -0.065 to +0.059 | 3/0/7 | 124–124 / 213–213 | 10–38 | | TESSERA v1 | loss | Control-mean adjustment | 10 | -0.018 to +0.064 | 4/0/6 | 124–124 / 213–213 | 3–23 | | TESSERA v1 | loss | Direct gradient boosting | 10 | -0.124 to +0.208 | 6/0/4 | 124–124 / 213–213 | 64–92 | Magnitude-score gain true-positive ranges across these same partitions are AlphaEarth V1 mosaic centre: 0–1; AlphaEarth V1 mosaic patch: 0–0; TESSERA v1 centre: 0–0; TESSERA v1 patch: 0–0. The nonzero AlphaEarth centre condition is seed 14015: 1/24 true positives and 17/85 false positives. This bounds the original zero-count observation; the original partition is not replaced. All partition-specific counts, operating points, conditional intervals, cell memberships, fitting counts and patch-support counterparts remain in analysis/partition-sensitivity-v24. This check examines one fitting/calibration choice; it does not estimate independent training-data, encoder or future-year variance. The original results remain primary. ## Revision history Revised in response to review-014-1: narrow the title/question to CONUS one-year tests; shorten the abstract; add primary AUROC, exact-fold reference/cohort variation, ranking similarity and year/region denominators; foreground the gain-ranking inversion; qualify AlphaEarth training-target and production-version dependence; move secondary/exploratory checks to appendices. All original scores, thresholds, failures and first-draft snapshots are retained. New saved-data diagnostics are exploratory and fixed after review in protocol/review-analysis-v1.md. Point-by-point decisions, access limits and change records are in review/response-1.md.