CAP-DBNet: A Class-Adaptive Prototype-Guided Dual-Branch Collaborative Network for Remote Sensing Image Semantic Segmentation

High-resolution remote sensing image semantic segmentation faces challenges from large-scale variation, fine boundary preservation, and class imbalance. This paper proposes CAP-DBNet, a compact framework that combines a RepViT M2.3 backbone with a global–local collaborative decoder (GLCD), a class-adaptive hybrid loss (CAHL), and a class-adaptive prototype-guided calibration module (CAPC). CAP-DBNet integrates class-center global and window-local modeling through bidirectional feature exchange, mini-batch class-frequency-aware supervision, and prototype-based residual calibration. Rather than introducing an isolated new attention or prototype mechanism, this work focuses on the coordinated design of these components within an efficient segmentation framework. In five repeated experiments with different random seeds, CAP-DBNet achieves average mIoUs of 85.16%, 88.32%, and 55.75% on Vaihingen, Potsdam, and LoveDA, respectively, improving over the same-backbone LOGCAN++ baseline by 1.43, 1.24, and 1.44 percentage points. Compared with D2LS under our evaluation protocol, CAP-DBNet uses 3.55× fewer parameters and requires 8.35× fewer floating-point operations (FLOPs). These results demonstrate an accuracy-efficiency-oriented global–local and class-adaptive framework for remote sensing semantic segmentation across three public benchmarks.