Abstract
Remote Sensing Change Detection (RSCD) is often hindered by environmental noise and the difficulty of maintaining precise object boundaries. This paper introduces STNet, a Siamese architecture that integrates a Temporal Feature Fusion (TFF) module and a Spatial Feature Fusion (SFF) module. STNet utilizes a cross-temporal gating mechanism to suppress irrelevant spectral variations and a cross-scale attention mechanism to refine spatial details. Experimental results on WHU, LEVIR-CD, and CLCD datasets demonstrate that STNet achieves state-of-the-art performance with significantly lower computational complexity (9.61 GFLOPs) compared to existing models.
Introduction
Change detection in bi-temporal satellite imagery is critical for monitoring urban expansion and land-use changes. However, two primary challenges persist: (1) "changes of non-interest," such as seasonal vegetation shifts, shadows, and lighting differences, which introduce spectral noise; and (2) the "boundary problem," where the sparse nature of changed objects makes it difficult to recover sharp edges after multiple downsampling layers. Traditional methods using simple feature subtraction or concatenation often fail to distinguish between semantic land-use changes and environmental variations.
Methodology
STNet employs a Siamese ResNet-18 backbone to extract multi-scale features from bi-temporal images $T$ and $T'$. The architecture is defined by two core fusion components:
Temporal Feature Fusion (TFF)
The TFF module addresses environmental noise through a cross-temporal gating mechanism. Instead of simple subtraction, it generates a coarse change map to weight the original features. This "gating" process prioritizes regions with significant semantic change while suppressing areas affected by seasonal or lighting variations.
Spatial Feature Fusion (SFF)
To preserve spatial precision, the SFF module implements a cross-scale attention mechanism. It leverages high-level semantic context from deeper layers to guide the fusion of low-level spatial features from shallower layers. This top-down guidance ensures that the model maintains sharp boundaries for small or sparse objects, such as building footprints.
The final stage uses a lightweight decoder with a Channel Attention Module (CAM) and a hybrid loss function (Focal Loss + Dice Loss) to address class imbalance.
Experiments and Results
STNet was evaluated on three benchmark datasets: WHU Building, LEVIR-CD, and CropLand Change Detection (CLCD).
Quantitative Performance
On the LEVIR-CD dataset, STNet achieved an F1-score of 90.52% and an IoU of 82.09%. Notably, the model is highly efficient, requiring only 9.61 GFLOPs. This is a substantial reduction compared to other high-performing models like SNUNet (54.88 GFLOPs) and IFNet (82.26 GFLOPs), making it suitable for large-scale deployment.
Qualitative Analysis
Visual comparisons indicate that STNet produces cleaner change maps with fewer false positives in regions with shadows or vegetation changes. The SFF module effectively maintains the geometric integrity of building corners and edges, outperforming models like SNUNet and ChangeFormer in boundary precision.
Conclusion
STNet provides a robust and efficient solution for remote sensing change detection. By combining cross-temporal gating and cross-scale attention, the model effectively filters environmental noise and recovers fine-grained spatial details. Future research will explore the model's applicability to heterogeneous data sources and its performance in semi-supervised learning environments.
Summary updated