Segmentation of Agricultural Fields on Microwave C-Band SAR Images

Abstract

The results of agricultural fields segmentation on microwave SAR images using several architectures of convolutional neural networks are presented. There are used IncFCN, MPResNet architectures, as well as U-Net and DeeplabV3+ modifications with ResNet34, ResNet50, Xception backbones. A study was carried out with various variations of loss functions and optimization algorithms. When training the selected architectures, the IncFCN network with the RMSprop learning algorithm and the Dice loss function showed the best result, and the Dice and F1 metrics reached 0.75 and 0.7, respectively. Based on the architecture showing the best values of the metrics, the calculation of the mRVI and NDVI vegetation indices for microwave and optical data, respectively, is given.