Advanced Fully Convolutional Networks for Agricultural Field Boundary Detection

Accurate spatial information of agricultural fields is important for providing actionable to farmers, managers, and policymakers. On the other hand, automated detection field boundaries a challenging task due their small size, irregular shape use mixed-cropping systems making vaguely defined. In this paper, we propose strategy boundary based on fully convolutional network architecture called ResU-Net. The benefits model are two-fold: first, residual units ease training deep networks. Second, rich skip connections within could facilitate propagation, allowing us design networks with fewer parameters but better performance in comparison traditional U-Net model. An extensive experimental analysis performed over whole Denmark using Sentinel-2 images comparing several ResU-Net algorithms. presented results show that has an average F1 score 0.90 Jaccard coefficient 0.80 0.88 0.77.