validation / comparison with ground observation / intercomparison

Jean-Baptiste Féret, Florian de Boissieu

2026-02-19

This tutorial illustrates how to perform validation of spectral diversity metrics based on ground information. Ground information should be associated with a vector layer including polygons. The size of the polygons should be compatible with ground information and with remote sensing information.

biodivMapR does not aim at producing diversity metrics with absolute accuracy:
the spectral diversity metrics produced with biodivMapR are not expected to match the absolute value of diversity metrics. Hence, when performing validation of diversity metrics with ground observations, users should favor correlation analyses, rather than using metrics such as RMSE between biodiversity indices measured from inventories and diversity metrics produced withbiodivMapR .

The choice of the number of clusters is directly influencing the range of the spectral $\alpha$ diversity metrics: higher number of clusters lead to higher maximum potential values of spectral richness, Shannon index and Hill index.

The function biodivMapR_opt_clusters can be used to identify the number of clusters maximizing the correlation between spectral diversity metrics and observations.

Ground information should not correspond to individual points:
biodivMapR produces diversity metrics based on information extracted from a group of pixels. $\alpha$ and $\beta$ spectral diversity metrics are computed from the distribution of cluster populations is each polygon. Therefore we recommend that each polygon defining an validation plots should correspond to at least 25-50 pixels of the image.

This code computes $\alpha$ and $\beta$ spectral diversity metrics over the footprint defined by polygons of a vector layer.

validation of biodivMapR using spectral features obtained from SPCA

The function get_diversity_from_plots can be performed using the biodivMapR spectral diversity models adjusted previously and saved in the Kmeans_info.RData and Beta_info.RData variables. The following code illustrates how to proceed, based on results obtained from previous tutorials. As we do not have ground information available for these sites, we will compare spectral diversity metrics obtained with the two sets of spectral features: selected components from a PCA, and a selection of spectral indices.

First, we randomly distribute square polygons with a surface matching with window_size.

# get extent of area of interest
input_rast <- terra::rast(rast_path)
extent_area <- get_raster_extent(input_rast = input_rast)
# sample data
samples <- sf::st_sample(x = sf::st_as_sf(extent_area), size = 5000,
                         force = TRUE)
# get resolution and adjust window around samples
raster_res <- terra::res(input_rast)
# apply buffer
bufferSize <- raster_res*window_size/2
samples <- terra::buffer(x = terra::vect(samples), width = bufferSize,
                         quadsegs = 8, capstyle  = 'square')

Then, spectral diversity is computed for these polygons based on the diversity mapped from PCA results.

# get spectral diversity from plots distributed over SPCA data
input_rast_pca <- terra::rast(PCA_Output$PCA_Files$PCA)
diversity_pca <- get_diversity_from_plots(input_rast = input_rast_pca, 
                                          validation_vect = samples, 
                                          input_mask = terra::rast(mask_path_PCA), 
                                          Kmeans_info = ab_info_SPCA$Kmeans_info, 
                                          Beta_info = ab_info_SPCA$Beta_info, 
                                          selected_bands = selected_bands, 
                                          alpha_metrics = c('richness', 'shannon', 'hill', 'simpson'), 
                                          Hill_order = 1)

Next, spectral diversity is computed for these polygons based on the diversity mapped from spectral indices.

# get spectral diversity from plots distributed over spectral index data
input_rast_si <- list('NDVI' = terra::rast(file.path(output_dir, 'spectral_indices_vrt/amazon_001_2024-08-23_NDVI.vrt')), 
                      'CR_SWIR' = terra::rast(file.path(output_dir, 'spectral_indices_vrt/amazon_001_2024-08-23_CR_SWIR.vrt')), 
                      'LAI_SAVI' = terra::rast(file.path(output_dir, 'spectral_indices_vrt/amazon_001_2024-08-23_LAI_SAVI.vrt')))
diversity_si <- get_diversity_from_plots(input_rast = input_rast_si, 
                                          validation_vect = samples, 
                                          input_mask = terra::rast(mask_path_PCA), 
                                          Kmeans_info = ab_info_SI$Kmeans_info, 
                                          Beta_info = ab_info_SI$Beta_info, 
                                          selected_bands = si_list, 
                                          alpha_metrics = c('richness', 'shannon', 'hill', 'simpson'), 
                                          Hill_order = 1)

validation biodivMapR using spectral indices

The validation can also be performed using the biodivMapR spectral diversity models adjusted on spectral indices. The following code illustrates how to proceed, based on results obtained from previous tutorials.

library(ggplot2)
library(RColorBrewer)
library(gridExtra)

# 1- list vector files corresponding to groups of validation plots
listShp <- as.list(list.files(path = output_dir_val, 
                              pattern = '.shp', full.names = T))

# 2- A SpatVectorCollection is created as validation plots correspond to multiple files
samplesVal <- lapply(listShp, terra::vect)
validation_vect <- terra::svc(samplesVal)

# 3- define SpatRaster for the set of spectral indices and corresponding mask
SI_rast <- lapply(SI_path,terra::rast)
mask_rast <- terra::rast(mask_path_SI)

# 4- perform validation over the footprint of each polygon
validation <- get_diversity_from_plots(input_rast = SI_rast,
                                       validation_vect = validation_vect,
                                       input_mask = mask_rast,
                                       Kmeans_info = ab_info_SI$Kmeans_info,
                                       Beta_info = ab_info_SI$Beta_info)

# 5- produce scatterplot using PCoA computed only from plots
filename <- file.path(output_dir_SI2,'BetaDiversity_Plots.png')
scatter_alphabeta(alpha = validation$specdiv$shannon_mean, 
                  PCoA_1 = validation$specdiv$BetaPlots_PCoA_1, 
                  PCoA_2 = validation$specdiv$BetaPlots_PCoA_2, 
                  PCoA_3 = validation$specdiv$BetaPlots_PCoA_3,
                  classes = validation$specdiv$source, 
                  filename = filename)

# produce scatterplot using PCoA computed from elements sampled across the full image
filename <- file.path(output_dir_SI2,'BetaDiversity_Full.png')
scatter_alphabeta(alpha = validation$specdiv$shannon_mean, 
                  PCoA_1 = validation$specdiv$BetaFull_PCoA_1, 
                  PCoA_2 = validation$specdiv$BetaFull_PCoA_2, 
                  PCoA_3 = validation$specdiv$BetaFull_PCoA_3,
                  classes = validation$specdiv$source, 
                  filename = filename)

The figures below illustrates the consistency in $\alpha$ and $\beta$ diversity metrics produced with biodivMapR over the same site with two different inputs (components selected from SPCA and spectral indices).

This intercomparison shows consistent estimation for $\alpha$ diversity metrics, with string correlation between shannon and hill indices obtained from the two types of data. However, the type of features used to compute spectral diversity influences Bray Curtis dissimilarity. The correlation of the Bray Curtis dissimilarity matrix remains moderate to strong.

Fig. 1. comparison of shannon, hill and bray curtis obtained from the two diversity maps obtained from SPCA and spectral indices.