<p>Class-incremental semantic segmentation methods for remote sensing images primarily rely on knowledge distillation to mitigate catastrophic forgetting. These methods overlook (1) the representation bias stemming from the conflict between cross-entropy loss and distillation loss, and (2) issues of interstep imbalance and interstep similarity that become particularly pronounced after multiple incremental learning steps, thus limiting their performance. To address these challenges, we propose MiR (mitigating representation bias), a novel framework that alleviates forgetting and facilitates learning new classes effectively. MiR replaces the traditional feature-classifier mode with the feature-SegToken interaction, leveraging implicit coarse-to-fine segmentation to mitigate representation bias. Furthermore, a weighting strategy is introduced to adaptively adjust cross-entropy and distillation losses, effectively tackling issues of interstep imbalance and interstep similarity. Extensive experiments conducted on the DeepGlobe, Potsdam, and Vaihingen datasets demonstrate the effectiveness of MiR in learning new knowledge while retaining old knowledge.</p>