The Segmentation of Unmanned Aerial Vehicle-Derived Aerial Photographs to Identify Anthropogenic Changes

The paper considers the issues of segmentation of aerial photographs obtained by Unmanned Aerial Vehicles to identify man-made changes. Neural networks are used for this purpose. Based on the orthophotoplan of the oil field area, a dataset was formed of about 4500 tiles measuring 512 × 512 pixels with 18 classes of man-made zones marked. The images were filtered, balanced and augmented, after which U‑Net, DeepLabv3+ and SegFormer models were trained on them with the training, test and validation sets divided into 70/15/15%. The best modification of U-Net showed an overall accuracy of 94.4% and mean Intersection over Union of 79.2%, while Intersection over Union values above 80% were obtained for key natural and man-made objects. DeepLabv3+ and SegFormer demonstrated comparable results (mean Intersection over Union about 74%) with better detail for large and rare classes. The proposed method ensures high accuracy and efficiency of analysis, which makes it promising for environmental monitoring.