Cross-Biome Transferability of SAR-based Flood Mapping with Random Forests

Fully automated, globally applicable flood-mapping systems must earn user trust, which in turn requires systematic testing across diverse environmental conditions to understand performance stability and a clear understanding of model transferability. While some recent studies have evaluated cross-site performance of flood mapping algorithms, the cross-biome transferability of Random Forest (RF) models for SAR-based flood delineation has not yet been thoroughly evaluated. In this study, we assess how well RF classifiers trained for binary flood detection generalize across biomes using primarily Synthetic Aperture Radar (SAR) data. Our feature stack comprises 14 variables, including 9 SAR-derived features (Sentinel-1 VV and VH backscatter and associated temporal-change metrics) which provide information on the flood-induced land surface changes and 4 contextual predictors such as land cover and topographic indices which influence radar backscatter and help to reduce as well as mitigate uncertainties. Experiments were conducted across 18 flood events distributed equally amongst 6 distinct biomes: (1) Deserts and Xeric Shrublands, (2) Tropical and Subtropical Moist Broadleaf Forests, (3) Temperate Broadleaf and Mixed Forests, (4) Temperate Coniferous Forests, (5) Mediterranean Forests, Woodlands and Scrub, (6) Temperate Grasslands, Savannas and Shrublands. Model transferability is evaluated using a two-level nested cross-validation approach. First, intra-biome performance is established through an inner 3-fold Leave-One-Group-Out Cross-Validation (LOGO-CV), in which models are trained on all but one site within a biome and evaluated on the held-out site iteratively. Second, inter-biome transferability is quantified using an outer 6-fold LOGO-CV, treating each biome as a distinct group. In this setup, models are trained on all biomes except one and evaluated on all sites of the held-out biome. Classification performance is assessed using Overall Accuracy (OA), F1-score, Precision, Recall, and Intersection over Union (IoU), with all experiments repeated across 10 independent iterations to capture model structural and sampling variability.

Preliminary results on select biomes show substantial variation in inter-biome transferability. Notably, in some cases, models transferred between biomes outperform those trained within the same biome. These findings highlight the need for comprehensive biome-level transferability assessments to better understand the capabilities and limitations of RF-based flood mapping under globally diverse conditions, ultimately supporting more transparent and trustworthy flood-mapping products for end users.