Accurate mapping of retrogressive thaw slumps (RTSs) in permafrost regions remains challenging due to their irregular morphology, blurred boundaries, and strong spatial correlation. This paper proposes a lightweight multi-level self-modulation (MLSM) module embedded into the UNet++ backbone to enhance non-local feature modeling for high-resolution image segmentation. The overall framework is built upon a UNet++ backbone with dense skip connections, where the proposed MLSM module adaptively fuses multi-scale contextual information to enhance feature coherence across spatially correlated regions. By incorporating a low-rank regularization term, MLSM dynamically modulates feature responses according to structural variations, allowing attention to adapt to spatially complex RTS regions. The integration of depth-wise convolution and channel recalibration further refines feature aggregation efficiency. Experimental evaluations on the Maxar dataset demonstrate that the proposed method achieves superior segmentation accuracy and smoother boundary delineation compared with existing models. The proposed framework provides a lightweight, robust, and computationally efficient solution for delineating irregular and morphologically complex RTSs.