A 3-D Fully Convolutional Network Approach for Land Cover Mapping Using Multitemporal Sentinel-1 SAR Data

Spaceborne temporal sequences of synthetic aperture radar (SAR) data have a definite advantage over multispectral data sequences in terms of continuity and regularity. Still, deep-learning (DL) applications in remote sensing have primarily focused on multispectral data. This work is focused instead on a novel 3-D DL architecture for SAR data sequences. The proposed approach utilizes a trained-from-scratch 3-D fully convolutional network (FCN) with a 3-D ResNet-50 as a backbone to classify ten land cover types using multitemporal Sentinel-1 SAR data. Experimental results show that this architecture provides a trained model that outperforms existing DL methods applied to the same SAR sequence in terms of overall accuracy (OA). In addition, the results using only SAR data provide very similar and consistent performances to those achievable using multispectral data. Accordingly, the proposed approach demonstrates the potential of SAR temporal sequences in land cover mapping using DL techniques.