4  Model Analysis

4.1 Load model and data

The trained model, calibration and validation data are loaded.

Code

# Load random forest model
rf_bor   <- readRDS(here::here("data/rf_for_pH0-10.rds"))
df_train <- readRDS(here::here("data/cal_for_ph0-10.rds"))
df_test  <- readRDS(here::here("data/val_for_ph0-10.rds"))

Next, we load a mask of the area over which the soil will be mapped. Our target area to predict over is defined in the file area_to_be_mapped.tif. Since we only want to predict on a given study area, the TIF file comes with a labeling of 0 for pixels that are outside the area of interest and 1 for pixels within the area of interest.

Code

# Load area to be predicted
raster_mask <- terra::rast(here::here("data-raw/geodata/study_area/area_to_be_mapped.tif"))

# Turn target raster into a dataframe, 1 px = 1 cell
df_mask <- as.data.frame(raster_mask, xy = TRUE)

# Filter only for area of interest
df_mask <- df_mask |> 
  dplyr::filter(area_to_be_mapped == 1)

# Display df
head(df_mask) |> 
  knitr::kable()

x	y	area_to_be_mapped
2587670	1219750	1
2587690	1219750	1
2587090	1219190	1
2587090	1219170	1
2587110	1219170	1
2587070	1219150	1

Next, we have to load the selected set of covariates as maps. These will be as the basis for spatial upscaling and provide the predictor values across space, fed into the trained model for predicting soil pH across space.

Get a list of all available covariate file names.

Code

files_covariates <- list.files(
  path = here::here("data-raw/geodata/covariates/"), 
  pattern = ".tif$",
  recursive = TRUE, 
  full.names = TRUE
  )

Note that the predictor rasters have to have the same resolution, extent, and coordinate reference system. This is the case as shown for two randomly picked examples.

Code

random_files <- sample(files_covariates, 2)
terra::rast(random_files[1])

class       : SpatRaster 
dimensions  : 986, 2428, 1  (nrow, ncol, nlyr)
resolution  : 20, 20  (x, y)
extent      : 2568140, 2616700, 1200740, 1220460  (xmin, xmax, ymin, ymax)
coord. ref. : CH1903+ / LV95 
source      : Se_curvplan2m_fmean_5c.tif 
name        : Se_curvplan2m_fmean_5c 
min value   :              -22.83849 
max value   :               29.64258 

Code

terra::rast(random_files[2])

class       : SpatRaster 
dimensions  : 986, 2428, 1  (nrow, ncol, nlyr)
resolution  : 20, 20  (x, y)
extent      : 2568140, 2616700, 1200740, 1220460  (xmin, xmax, ymin, ymax)
coord. ref. : CH1903+ / LV95 
source      : Se_curvplan50m.tif 
name        : Se_curvplan50m 
min value   :      -3.427344 
max value   :       3.426279 

Load the rasters for the selected predictor variables into a raster object (a “stack” of multiple rasters).

Code

# Filter that list only for the variables used in the RF
preds_selected <- names(rf_bor$forest$covariate.levels)
files_selected <- files_covariates[apply(sapply(X = preds_selected, 
                                            FUN = grepl, 
                                            files_covariates), 
                                     MARGIN =  1, 
                                     FUN = any)]

# Load all rasters as a stack
raster_covariates <- terra::rast(files_selected)

Convert the raster stack into a dataframe - the preferred format for model prediction.

Code

# Get coordinates for which we want data
df_locations <- df_mask |> 
  dplyr::select(x, y)

# Extract data from covariate raster stack for all gridcells in the raster
df_predict <- terra::extract(
  raster_covariates,   # The raster we want to extract from
  df_locations,        # A matrix of x and y values to extract for
  ID = FALSE           # To not add a default ID column to the output
  )

df_predict <- cbind(df_locations, df_predict) |> 
  tidyr::drop_na()  # Se_TWI2m has a small number of missing data

4.2 Model testing

4.2.1 Make predictions

To test our model for how well it predicts on data it has not used during model training, we first have to load the {ranger} package to load all functionalities to run a Random Forest with the predict() function. Alongside our model, we feed our validation data into the function and set its parallelization settings to use all but one of our computer’s cores.

Code

# Need to load {ranger} because ranger-object is used in predict()
library(ranger) 

# Make predictions for validation sites
prediction <- predict(
  rf_bor,           # RF model
  data = df_test,   # Predictor data
  num.threads = parallel::detectCores() - 1
  )

# Save predictions to validation df
df_test$pred <- prediction$predictions

4.2.2 Model metrics

Now that we have our predictions ready, we can extract standard metrics for a regression problem (see AGDS Chapter 8.2.2).

Code

# Calculate error
err <- df_test$ph.0.10 - df_test$pred

# Calculate bias
bias <- mean(err, na.rm = TRUE) |> round(2)

# Calculate RMSE
rmse <- sqrt(mean(err, na.rm = TRUE)) |> round(2)

# Calculate R2
r2 <- cor(df_test$ph.0.10, df_test$pred, method = "pearson")^2 |> round(2)

4.2.3 Metric plots

Code

df_test |> 
  ggplot2::ggplot(ggplot2::aes(x = pred, y = ph.0.10)) +
  ggplot2::geom_point() +
  ggplot2::geom_smooth(method = "lm",
                       color = "tomato") +
  ggplot2::theme_classic() +
  ggplot2::geom_abline(
    intercept = 0, 
    slope = 1, 
    linetype = "dotted") +
  ggplot2::ylim(5, 7.5) +
  ggplot2::xlim(5, 7.5) +
  ggplot2::labs(
    title = "Predicted vs. Observed soil pH 0-10 cm",
    subtitle = bquote(paste("Bias = ", .(bias), 
                            ", RMSE = ", .(rmse), 
                            ", R"^2, " = ", .(r2))),
    x = "Predicted pH",
    y = "Observed pH"
  )

Figure 4.1: Comparison of observed versus predicted values for top soil pH using a simple Random Forest model.

The plot shows that our model explains about half of the observed variation in soil pH. Yet, we can also see that the model tends to overestimate low pH values. Anyways, let’s move ahead.

4.3 Create prediction maps

The fitted and tested model can now be used for spatially upscaling - creating a map of top soil pH values across our study area. For this, we again make predictions with our Random Forest model but we use our covariates dataframe for the study area, instead of only at the sampling locations as done above.

Code

# Make predictions using the RF model
prediction <- predict(
  rf_bor,              # RF model
  data = df_predict,   
  num.threads = parallel::detectCores() - 1)

# Attach predictions to dataframe and round them
df_predict$prediction <- prediction$predictions

Code

# Extract dataframe with coordinates and predictions
df_map <- df_predict |>
  dplyr::select(x, y, prediction)

# Turn dataframe into a raster
raster_pred <- terra::rast(
  df_map,                  # Table to be transformed
  crs = "+init=epsg:2056", # Swiss coordinate system
  extent = terra::ext(raster_covariates) # Prescribe same extent as predictor rasters
  )

Code

# Let's have a look at our predictions!
# To have some more flexibility, we can plot this in the ggplot-style as such:
ggplot2::ggplot() +
  tidyterra::geom_spatraster(data = raster_pred) +
  ggplot2::scale_fill_viridis_c(
    na.value = NA,
    option = "viridis",
    name = "pH"
    ) +
  ggplot2::theme_classic() +
  ggplot2::scale_x_continuous(expand = c(0, 0)) +
  ggplot2::scale_y_continuous(expand = c(0, 0)) +
  ggplot2::labs(title = "Predicted soil pH (0 - 10cm)")

Figure 4.2: Predicted map of top soil pH using a simple Random Forest model.

Interesting, we see that in this study area, there is a tendency of having more acidic soils towards the south west and more basic soils towards the north east.

Let’s write the predicted top soil pH raster into a GeoTIFF file.

Code

# Save raster as .tif file
terra::writeRaster(
  raster_pred,
  "../data/ra_predicted_ph0-10.tif",
  datatype = "FLT4S",  # FLT4S for floats, INT1U for integers (smaller file)
  filetype = "GTiff",  # GeoTiff format
  overwrite = TRUE     # Overwrite existing file
)

That’s it.