GIS-Integrated Semi-Supervised U-Net for Automated Spatiotemporal Detection and Visualization of Land Encroachment in Protected Areas Using Remote Sensing Imagery

With such illegal land use activities as gradually increasing within the restricted protected areas of the United States, it is urgent to carry out the efficient monitoring and management support for encroachment through remote sensing images and intelligent analysis. In this work, we develop a Semi-supervised Enhanced U-Net (SS-EU-Net) for remote sensing images in this article, with an ability to identify land encroachment areas, as well as to visualize spatially and analyze its time evolution behavior by means of the geographic information system (GIS). Inherited from the conventional U-Net, the model introduces two critical improvements: first, the method integrates self-supervised pre-training scheme and pseudo-label generation mechanism to improve the feature learning capability of the model on the unlabeled satellite images. Multi-scale attention fusion module was proposed to greatly enhance the segmentation performance on complex edge of ground objects. The model takes a remote sensing image with geographical coordinates as an input and then generates a pixel-level mask of encroachment based on GIS layer through coordinate registration, which accomplishes the refined monitoring for the temporal and spatial variations of encroachment as well as helps for the decision support. Experimental results demonstrate that SS-EU-Net enhances the IoU and F1 by 5.3% and 4.7% in remote sensing image data set of typical protected areas in USA than existing methods.