雷达-光学特征融合与自监督学习驱动的湿地动态监测OA

Addressing the challenges of cloud interference and multi-source data fusion in wetland monitoring in the Pearl River Delta,we pro-posed a spatio-temporal enhanced self-supervised attention network model(STE-SAN)integrating radar-optical spatio-temporal features.We inte-grated Sentinel-1 SAR and Landsat images to construct a spatio-temporal registered dataset,combined 3D convolution with bidirectional LSTM to extract multi-scale spatio-temporal features,and designed a polarization-sensitive attention mechanism to enhance the identification of wetland edge heterogeneity.The results show that the overall classification accuracy of STE-SAN reaches 91.7%(Kappa=0.88),which is 8.2 to 15.7 per-centage points higher than that of random forest and U-Net.It performs outstandingly in cloudy scenarios(F1=85.4%)and edge detection(IoU=0.81),and the annotation requirement is only 5%of that of traditional methods.The monitoring results indicated that from 2000 to 2020,the natural wetlands in the Pearl River Delta decreased by 15.9%(while the artificial wetlands increased by 111.5%),and 83%of the coastal wet-lands were encroached by aquaculture ponds,leading to the degradation of ecological functions.This model provides a high-precision and low-cost solution for wetland dynamic monitoring and conservation decision-making in highly urbanized areas.