Performance Analysis of Super-Resolution-Inspired Deep Learning Methods for Forest Segmentation
This talk focuses on the application of deep learning methods inspired by super-resolution techniques to high-resolution vegetation segmentation. The study uses Sentinel-2 satellite imagery at a spatial resolution of 10 m and high-resolution reference masks derived from LiDAR data. The presentation covers the dataset preparation process and compares architectures employing spatial detail reconstruction mechanisms based on residual and dense connections in various configurations. These models are designed to directly predict high-resolution segmentation masks without requiring an intermediate reconstruction of satellite imagery. The proposed approaches are evaluated against baseline models in terms of pixel-wise segmentation performance and the accuracy of vegetation boundary delineation.