DASF-Net: A Despeckling-Attentive SAR-Optical Fusion Network for Coastal Land Cover Mapping

Abstract

Rapid and often unplanned growth along India's coast makes timely land-use and land-cover (LULC) monitoring essential, yet the optical satellite imagery normally used for this task is frequently obscured by cloud during the monsoon over cities such as Mumbai. Synthetic Aperture Radar (SAR) acquires data throughout the year irrespective of cloud, but its weak spectral discrimination limits its use on its own. This paper studies whether fusing Sentinel-1 SAR with Sentinel-2 optical imagery can overcome these individual weaknesses, and proposes DASF-Net(Despeckling-Attentive SAR-Optical Fusion Network). The model places a learned residual despeckling stage on the SAR channels and feeds an eleven-channel SAR-optical tensor into a U-Net encoder-decoder that applies Convolutional Block Attention (CBAM) at its decoder stages and an Atrous Spatial Pyramid Pooling(ASPP) bottleneck for multi-scale context, while a boundary-aware auxiliary head sharpens class transitions. The inputs are built from dry-season composites for 2018 and 2024 and supervised with Esri 10m land-cover labels. On a spatially held-out test region, DASF-Net reaches an overall accuracy of 90.5%, a mean IoU of 0.577 and a weighted F1 of 0.910, surpassing a Random Forest classifier and U-Net and DeepLabV3+ baselines trained on the same data. Applying the model across the two years reveals an increase of about 7.6% in built-up area, dominated by the conversion of vegetation to built-up land.