Agricultural Fields Segmentation on Satellite Images Using Convolutional Neural Networks

Agricultural fields segmentation algorithms in satellite images are presented. Three convolutional neural networks were developed: U-Net with ResNet-34 and SE-ResNeXt-50 backbones and Deeplabv3+ with Xception backbone. All backbones were pretrained in Imagenet database. Training and testing of algorithms was carried out on NVIDIA DGX-I supercomputer using a dataset of high-resolution images from the Sentine1-2 satellite. The value of F1 and Dice were 0.706 and 0.942 for U-Net with SE-ResNeXt-50 backbone. Test results confirm high accuracy in determining the boundaries of agricultural fields by proposed algorithm. © 2021 IEEE.