DARES: Rényi-Entropy Domain Adaptation for Forest Deforestation Segmentation

Semantic segmentation of multispectral satellite imagery is central to large-scale tropical deforestation monitoring, yet cross-region deployment suffers from severe accuracy degradation due to radiometric, atmospheric, and geographic covariate shifts. We introduce dares (Domain Adaptation via Rényi Entropy for Semantic Segmentation), an unsupervised domain adaptation framework that transfers deforestation-mapping models across biomes without target annotations for training. Its core mechanism performs class-conditional Rényi-2 (α=2) alignment on the deepest feature grid together with decoupled confidence-weighted Rényi-2 entropy minimization on target predictions, stabilized against a degenerate collapse as detailed in Methods. Evaluated on an ecologically critical transfer task from the Brazilian Amazon to the Colombian deforestation arc using Sentinel-2 MSI data, dares attains 64.65% target-test mIoU (78.43% DICE) under moderate radiometric attenuation, numerically outperforming Source-Only (44.24%), ADVENT (50.24%), and CBST (48.92%) while maintaining 76.75% source-domain discriminability. Across a Low→Medium→High stress trajectory, dares follows a non-monotonic path from 75.42%→64.65%→69.54%, a 5.9-point Low-to-High change with a Medium dip and a High rebound. It is not the best at Low, where ADVENT (85.78%) and Source-Only (78.93%) exceed its 75.42%, while its advantage opens from Medium (+14.4 points over ADVENT, +20.4 over Source-Only) at the cost of the lowest source-test mIoU (76.75%). Component analysis indicates that decoupled entropy minimization drives target adaptation (−17.2 points if removed), whereas alignment, anti-collapse floors, and repulsion act as source-stabilizing regularizers rather than target maximizers.