DARES: Rényi-Entropy Domain Adaptation for Forest Deforestation Segmentation

Source abstract

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, at-mospheric, 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. DARES generalizes matrix-based conditional Rényi-2 (α = 2) divergence to dense prediction by aligning class-conditional representations extracted from the deepest feature grid under a uniform quota sampler. To guarantee numerical stability and prevent degenerate Dirac collapse, the latent alignment is reinforced by balanced trace-normalized mix-tures, asymmetric stop-gradient anchoring, spectral entropy floors, margin-hinged inter-class repulsion, and a trust-region gradient scheduler, coupled with a decoupled confidence-weighted Rényi-2 entropy minimization on target predictions. Evaluated on an ecologically critical transfer task from the Brazilian Amazon to the Colombian deforestation arc using Sentinel-2 MSI data, DARES attains 63.70% target-test mIoU (77.75% DICE) under moderate radiometric attenuation, significantly outperforming Source-Only (35.41%), ADVENT (45.71%), and CBST (54.71%), while maintaining 81.75% source-domain discriminabil-ity. Across a progressive stress trajectory, DARES exhibits graceful degradation (bounded to 17.14 mIoU points) and prevents the catastrophic collapse observed in adversarial and self-training baselines under severe attenuation. Component analysis indicates that while decoupled entropy minimization drives spatial target polarization, the latent safeguards establish an essential Pareto-regularization envelope that preserves structural manifold stability.