Enhancing Few-Shot Semantic Segmentation in Remote Sensing Through Magnitude-Based Pruning

Abstract

Few-Shot Semantic Segmentation (FSSS) in remote sensing faces significant challenges owing to the limited availability of labeled data and the complexity of high-resolution imagery. To address these challenges, we propose a novel framework that integrates magnitude-based pruning with Cross-Matching and Self-Matching modules. By systematically pruning 30% of the redundant weights from the backbone network, we enhanced the feature extraction and segmentation accuracy while maintaining the model efficiency. Experimental evaluations of the DLRSD-5i and ISAID-5i datasets demonstrated the effectiveness of the proposed method. On DLRSD-5i, the pruned SCCNet achieved a mean mIoU improvement of +9.40 (1-shot) and +6.46 (5-shot) compared to the baseline, outperforming state-of-the-art (SOTA) models. Similarly, on ISAID-5i, the pruned ResNet-101 surpassed SOTA by +1.18 (1-shot) and +0.69 (5-shot) in mean mIoU. These results validate the effectiveness of pruning in optimizing the baseline model for FSSS tasks, thereby enhancing its ability to generalize and accurately segment complex remote-sensing imagery.

Keywords: Few-shot learning, Neural network pruning, Semantic segmentation, Remote sensing, Backbone networks, Feature Extraction