Using a semantic edge-aware multi-task neural network to delineate agricultural parcels from remote sensing images

Abstract

This paper presents a semantic edge-aware multi-task neural network (SEANet) to obtain closed boundaries when delineating agricultural parcels from remote sensing images. It derives closed boundaries from remote sensing images and improves conventional semantic segmentation methods for the extraction of small and irregular agricultural parcels. SEANet integrates three correlated tasks: mask prediction, edge prediction, and distance map estimation. Related features learned from these tasks improve the generalizability of the network. We regard boundary extraction as an edge detection task and extract rich semantic edge features at multiple levels to improve the geometric accuracy of parcel delineation. Moreover, we develop a new multi-task loss that considers the uncertainty of different tasks. We conducted experiments on three high-resolution Gaofen-2 images in Shandong, Xinjiang, and Sichuan provinces, China, and on two medium-resolution Sentinel-2 images from Denmark and the Netherlands. Results showed that our method produced a better layout of agricultural parcels, with higher attribute and geometric accuracy than the existing ResUNet, ResUNet-a, R2UNet, and BsiNet methods on the Shandong and Denmark datasets. The total extraction errors of the parcels produced by our method were 0.214, 0.127, 0.176, 0.211, and 0.184 for the five datasets, respectively. Our method also obtains closed boundaries by one single segmentation, leading to superiority as compared with existing multi-task networks. We showed that it could be applied to images with different spatial resolutions for parcel delineation. Finally, our method trained on the Xinjiang dataset could be successfully transferred to the Shandong dataset with different dates and landscapes. Similarly, we obtained satisfactory results when transferring from the Denmark dataset to the Netherlands dataset. We conclude that SEANet is an accurate, robust, and transferable method for various areas and different remote sensing images. The codes of our model are available at https://github.com/long123524/SEANet_torch.