How Accurate Is Your Conservation Map?

Divide the study area into spatial blocks to ensure training and validation data independence. All sampling and cross-validation is performed at the block level. Gold-highlighted blocks in the example below show out-of-bag (OOB) blocks for one bootstrap iteration (~36.8% of total).

Draw B bootstrap samples of spatial blocks (with replacement). For each replicate, train a Random Forest on drawn blocks and validate on the ~36.8% of blocks not drawn (out-of-bag approach). This produces B independent accuracy estimates from which distributions and confidence intervals are derived.

Distributions of accuracy metrics across all B bootstrap replicates. 95% confidence intervals indicate the precision with which accuracy is known. Dashed red line indicates the 85% accuracy threshold. Dashed red line indicates threshold values (R² > 0.8, rel. RMSE < 0.2).

Error-matrix-corrected area estimates with 95% confidence intervals (Olofsson et al. approach). Each bootstrap replicate produces an independent error matrix used to adjust raw pixel counts, accounting for classification errors. Bias-corrected total biomass estimates with 95% confidence intervals. Mean OOB residuals from each bootstrap are used to correct population sums.