Differentiable energy-guided modulation for remote sensing image segmentation

Precise segmentation of remote sensing images is challenged by multi-scale object distributions, blurred boundaries, and class inconsistencies. We propose Differentiable Energy-Guided Modulation (DEGM), a lightweight framework that integrates explicit energy-based priors with data-driven refinement to enhance semantic representation. Local feature energy is extracted from shallow and deep encoder layers to generate a soft differentiable prior mask, which is refined through nonlinear mapping into an adaptive spatial mask. This fused mask guides the network toward discriminative regions and sharpens object boundaries. A hybrid loss combining cross-entropy and Dice objectives jointly optimizes pixel accuracy and boundary precision. Incorporated into U-Net with negligible overhead, DEGM remains fully end-to-end trainable and easily adaptable to other encoder–decoder backbones. Experiments on the WHDLD and DLRSD datasets show that our method outperforms U-Net, achieving a mean F1-score of 75.63% on WHDLD and 84.54% on DLRSD, with improvements of 1.64% and 2.12%, respectively, while maintaining computational efficiency.