Thanks to advances in deep learning for image processing and pattern recognition, remote sensing data classification has progressed significantly in the last few years. Pre-trained deep learning models promise fast and automated analysis of spatial data, but their applicability in new spatial and content contexts often remains unexplored. In this study, we evaluate the performance of four pre-trained deep learning models integrated into ArcGIS Pro on Slovenia’s national digital orthophotos (DOF) from 2018 and 2021. The models tested were Building Footprint Extraction Model – USA, Segment Anything Model (SAM), Tree Segmentation Model and Road Extraction – North America Model. The analysis was conducted for two spatially and contextually different areas, one urban (the settlement Brdo in Ljubljana) and one rural (the municipality of Ig). Results demonstrate that tuning key parameters can improve model performance, but that inherent limitations persist when applying these models to local orthophotos. The highest accuracy was achieved for building footprint extraction, whereas the tree and road segmentation models exhibited higher rates of false detections. Overall, pre-trained models offer a valuable starting point for processing very high resolution (VHR) imagery, but attaining higher precision requires additional fine-tuning and re-training on locally representative data.
|
