---
title: 'Practical 5: Estimating variance'
author: David L Miller
css: custom.css

---

Here is a "solution" for practical 5. As with any data analysis, there is no correct answer, but this shows how I would approach this analysis. The analysis here is conditional on selecting a detection function and DSM in the previous exercises; I've shown a variety of models selected in the previous solutions to show the differences between models.

Much of the text below is as in the exercise itself, so it should be relatively easy to navigate.

<div class="boxout">
Additional text and code is highlighted using boxes like this.
</div>

Now we've fitted some models and estimated abundance, we can estimate the variance associated with the abundance estimate (and plot it).

# Load packages and data

```{r load-packages}
library(dsm)
library(raster)
library(ggplot2)
library(viridis)
library(plyr)
library(knitr)
library(rgdal)
```

Load the models and prediction grid:
```{r load-models}
load("dsms.RData")
load("dsms-xy.RData")
load("predgrid.RData")
```

# Estimation of variance

Depending on the model response (count or Horvitz-Thompson) we can use either `dsm.var.prop` or `dsm.var.gam`, respectively. `dsm_nb_xy_ms` doesn't include any covariates at the observer level in the detection function, so we can use the variance propagation method and estimate the uncertainty in detection function parameters in one step.

```{r varest}
# need to remove the NAs as we did when plotting
predgrid_var <- predgrid[!is.na(predgrid$Depth),]
# now estimate variance
var_nb_xy_ms <- dsm.var.prop(dsm_nb_xy_ms, predgrid_var, off.set=predgrid_var$off.set)
```
To summarise the results of this variance estimate:
```{r varest-summary}
summary(var_nb_xy_ms)
```

Try this out for some of the other models you've saved. Remember to use `dsm.var.gam` when there are covariates in the detection function and `dsm.var.prop` when there aren't.

<div class="boxout">
We'll skip this and go straight to summarising all our models so far in the next section...
</div>

# Summarise multiple models

We can again summarise all the models, as we did with the DSMs and detection functions, now including the variance:

```{r summarise-var}
summarize_dsm_var <- function(model, predgrid){

  summ <- summary(model)
  
  vp <- summary(dsm.var.prop(model, predgrid, off.set=predgrid$off.set))
  unconditional.cv.square <- vp$cv^2
  asymp.ci.c.term <- exp(1.96*sqrt(log(1+unconditional.cv.square)))
  asymp.tot <- c(vp$pred.est / asymp.ci.c.term,
                 vp$pred.est,
                 vp$pred.est * asymp.ci.c.term)

  data.frame(response = model$family$family,
             terms    = paste(rownames(summ$s.table), collapse=", "),
             AIC      = AIC(model),
             REML     = model$gcv.ubre,
             "Deviance_explained" = paste0(round(summ$dev.expl*100,2),"%"),
             "lower_CI" = round(asymp.tot[1],2),
             "Nhat" = round(asymp.tot[2],2),
             "upper_CI" = round(asymp.tot[3],2)
             )
}
```
<div class="boxout">
```{r var-table}
# make a list of models
model_list <- list(dsm_nb_xy, dsm_nb_x_y, dsm_nb_xy_ms, dsm_nb_x_y_ms,
                           dsm_tw_xy, dsm_tw_x_y, dsm_tw_xy_ms, dsm_tw_x_y_ms)
# give the list names for the models, so we can identify them later
names(model_list) <- c("dsm_nb_xy", "dsm_nb_x_y", "dsm_nb_xy_ms", "dsm_nb_x_y_ms",
                           "dsm_tw_xy", "dsm_tw_x_y", "dsm_tw_xy_ms", "dsm_tw_x_y_ms")
per_model_var <- ldply(model_list, summarize_dsm_var, predgrid=predgrid_var)
```
</div>

```{r print-table}
kable(per_model_var)
```


# Plotting

We can plot a map of the coefficient of variation, but we first need to estimate the variance per prediction cell, rather than over the whole area. This calculation takes a while!

```{r per-cell-var}
# use the split function to make each row of the predictiond data.frame into
# an element of a list
predgrid_var_split <- split(predgrid_var, 1:nrow(predgrid_var))
var_split_nb_xy_ms <- dsm.var.prop(dsm_nb_xy_ms, predgrid_var_split, off.set=predgrid_var$off.set)
```
Now we have the per-cell coefficients of variation, we assign that to a column of the prediction grid data and plot it as usual:

```{r varest-map-obs}
predgrid_var_map <- predgrid_var
cv <- sqrt(var_split_nb_xy_ms$pred.var)/unlist(var_split_nb_xy_ms$pred)
predgrid_var_map$CV <- cv
p <- ggplot(predgrid_var_map) +
       geom_tile(aes(x=x, y=y, fill=CV, width=10*1000, height=10*1000)) +
       scale_fill_viridis() +
       coord_equal() +
       geom_point(aes(x=x,y=y, size=count), data=dsm_nb_xy_ms$data[dsm_nb_xy_ms$data$count>0,])
print(p)
```

Note that here we overplot the segments where sperm whales were observed (and scale the size of the point according to the number observed), using `geom_point()`.

We can also overplot the effort, which can be a useful way to see what the cause of uncertainty is. Though it may not only be caused by lack of effort but also covariate coverage, this can be useful to see.

First we need to load the segment data from the `gdb`
```{r load-effort}
tracks <- readOGR("Analysis.gdb", "Segments")
tracks <- fortify(tracks)
```
We can then just add this to the plot object we have built so far (with `+`), but this looks a bit messy with the observations, so let's start from scratch:
```{r varest-map-obs-effort}
p <- ggplot(predgrid_var_map) +
       geom_tile(aes(x=x, y=y, fill=CV, width=10*1000, height=10*1000)) +
       scale_fill_viridis() +
       coord_equal() +
       geom_path(aes(x=long,y=lat, group=group), data=tracks)
print(p)
```

Try this with the other models you fitted and see what the differences are between the maps of coefficient of variation.

<div class="boxout">
We can use a similar technique as we did in the prediction exercises to get coefficient of variation maps for all of the models...

**Note this can take a long time!**

```{r predsp-allgen, fig.width=15, fig.height=7}
# make a function that calculates the CV, adds them to a column named CV
# and adds a column called "model" that stores the model name, then returns the
# data.frame.
make_cv_dat <- function(model_name, predgrid_split, predgrid){
  # we use get() here to grab the object with the name of its argument
  var_obj <- dsm.var.prop(get(model_name), predgrid_split, off.set=predgrid$off.set)
  predgrid[["CV"]] <- sqrt(var_obj$pred.var)/unlist(var_obj$pred)
  predgrid[["model"]] <- model_name
  return(predgrid)
} 

# load plyr and apply to a list of the names of the models, make_cv_dat returns
# a data.frame (hence this is an "ld" function: list->data.frame) that it then binds
# together
library(plyr)
big_cv <- ldply(list("dsm_nb_xy", "dsm_nb_x_y", "dsm_nb_xy_ms", "dsm_nb_x_y_ms",
                     "dsm_tw_xy", "dsm_tw_x_y", "dsm_tw_xy_ms", "dsm_tw_x_y_ms"),
                make_cv_dat, predgrid_split=predgrid_var_split, predgrid=predgrid_var_map)
```

```{r predsp-allplot}
p <- ggplot(big_cv) +
       geom_tile(aes(x=x, y=y, fill=CV, width=10*1000, height=10*1000)) +
       scale_fill_viridis() +
       coord_equal() +
       facet_wrap(~model, nrow=2) +
       geom_path(aes(x=long,y=lat, group=group), data=tracks)
print(p)
```

One issue with producing this kind of plot is that `ggplot2` will use a common legend, meaning if there are some cells with high values, this can throw out the rest of the scale.

As suggested in the lectures, we'll now use `cut()` to create a categorised version of the CV and plot that instead:


```{r predsp-allplot-cut}
big_cv$CV_cut <- cut(big_cv$CV, c(seq(0,2,0.2),3:6))
p <- ggplot(big_cv) +
       geom_tile(aes(x=x, y=y, fill=CV_cut, width=10*1000, height=10*1000)) +
       scale_fill_viridis(discrete=TRUE) +
       coord_equal() +
       facet_wrap(~model, nrow=2) +
       geom_path(aes(x=long,y=lat, group=group), data=tracks)
print(p)
```

This is a bit easier to read.

</div>

# Save the uncertainty maps to raster files

As with the predictions, we'd like to save our uncertainty estimates to a raster layer so we can plot them in ArcGIS. Again, this involves a bit of messing about with the data format before we can save.

```{r savecv-raster}
# setup the storage for the cvs
cv_raster <- raster(predictorStack)
# we removed the NA values to make the predictions and the raster needs them
# so make a vector of NAs, and insert the CV values...
cv_na <- rep(NA, nrow(predgrid))
cv_na[!is.na(predgrid$Depth)] <- cv
# put the values in, making sure they are numeric first
cv_raster <- setValues(cv_raster, cv_na)
# name the new, last, layer in the stack
names(cv_raster) <- "CV_nb_xy"
```
We can then save that object to disk as a raster file:
```{r write-raster}
writeRaster(cv_raster, "cv_raster.img", datatype="FLT4S")
```


# Extra credit

- `dsm.var.prop` and `dsm.var.gam` can accept arbitrary splits in the data, not just whole areas or cells. Make a `list` with two elements: one a `data.frame` of all the cells with $y>0$ and one with $y\leq 0$. Estimate the variance for these regions. Note that you'll need to sum the offsets for each area to get the correct value to supply to `off.set=...`.


<div class="boxout">
Trying a North-South split about $y=0$...

```{r split-var}
# create a data frame
predgrid_var_ns <- list()
predgrid_var_ns[["North"]] <- predgrid_var[predgrid_var$y>0, ]
predgrid_var_ns[["South"]] <- predgrid_var[predgrid_var$y<=0, ]

# get the offsets
ns_offsets <- c(sum(predgrid_var[predgrid_var$y<=0, ]),
                sum(predgrid_var[predgrid_var$y>0, ]))

# calculate the variance per region
var_split_ns <- dsm.var.prop(dsm_nb_xy_ms, predgrid_var_ns, off.set=ns_offsets)
```

We can then pull out the CVs:

```{r split-cv}
sqrt(var_split_ns$pred.var)/unlist(var_split_ns$pred)
```

So the northern region has a lower CV than the southern region, which we expect given the distribution of effort.
</div>