LFANet: Low-Altitude Remote Sensing Change Detection With Level-Wise Feature Fusion and Multi-Scale Aggregation for Land-Use Monitoring

With the rapid growth of low-altitude remote sensing for Earth observation, change detection faces many challenges, such as class imbalance, variations in illumination and imaging conditions, and false changes caused by geometric distortions. To address these issues, we propose a Level-wise Feature Fusion and Multi-Scale Aggregation Network (LFANet). Specifically, LFANet includes a hierarchical feature fusion module that exploits the spatiotemporal differences between bi-temporal features and integrates them across layers with a weighting strategy. The resulting hierarchical difference representation provides fine spatiotemporal context for subsequent change discrimination. Furthermore, building on multi-scale feature interactions, we design a lightweight multi-scale aggregation module that uses feature mapping and multi-scale fusion to improve cross-scale alignment and effectively suppress non-differential noise. Experimental results show that the proposed LFANet achieves better performance than mainstream methods on several public benchmarks, demonstrating its potential for real-world remote sensing change detection.