A Decoder-Enhanced Network With Multiscale Aggregation for Multiclass Remote Sensing Change Detection

Multi-class land-cover change detection from bi temporal remote sensing imagery has become a critical task for large-scale Earth observation. Nevertheless, prevailing segmentation-based change detectors remain prone to false alarms and missed detections, especially along fine-grained boundaries and in complex backgrounds where unchanged regions are easily confused with particular change classes. To mitigate these issues, we propose a decoder-enhanced UNet style segmentation framework (USNet) to improve change-class discrimination by jointly reinforcing semantic representation, boundary/detail reconstruction, and computational efficiency. Specifically, USNet strengthens the bottleneck to enrich high level contextual semantics while adopting lightweight convo lutional blocks to constrain decoding complexity. Moreover, spatial-and-channel squeeze¨Cexcitation (SCSE) recalibration is incorporated into the up-sampling stages to suppress background interference and facilitate accurate boundary recovery. Extensive experiments on the SECOND dataset demonstrate that USNet consistently outperforms representative segmentation baselines. In particular, it achieves 72.69% mIoU, 49.35% F1-score, and 88.97% accuracy, surpassing strong competitors such as FCN, UNet, PSPNet, and DeepLabV3+ across all evaluation metrics. These results confirm the effectiveness of USNet for robust multi class land-cover change detection in challenging remote sensing scenarios.