Comparing UNet and HRNet for Geomorphic Feature Extraction from Digital Terrain Model-Derived Land Surface Parameters

Behnam Solouki, Aaron E Maxwell 

Automated extraction of geomorphic features from high spatial resolution terrain data is important for environmental monitoring, geohazard assessment, and landscape change analysis. This study compares two convolutional neural network architectures, UNet and High-Resolution Network (HRNet), for semantic segmentation of mine benches and sinkholes from lidar-derived digital terrain models and land surface parameters. Mine benches represent anthropogenic, elongated, stepped landforms associated with historic surface mining, whereas sinkholes are natural, compact depressions common in karst terrain. Three terrain-derived input layers were used: square-root-transformed slope, broad-scale topographic position index, and fine-scale annulus topographic position index. Models were trained using five training sample sizes, including 50, 100, 250, 500, and 1,000 chips, and evaluated using overall accuracy, precision, recall, and F1-score. Results suggest that both architectures achieved broadly comparable segmentation performance, with F1-scores generally improving as training sample size increased. HRNet showed a practically notable advantage for sinkholes at the smallest sample size, but this advantage diminished as more training data were provided. For mine benches, UNet and HRNet performed similarly across most sample sizes, with neither architecture showing a consistent advantage. Although HRNet had substantially greater computational complexity, requiring approximately 12.7 times more trainable parameters and about three times more FLOPs and MACs than UNet, this additional cost did not produce a corresponding accuracy gain. These findings indicate that, for terrain-only geomorphic feature extraction, baseline UNet offers a highly efficient and competitive solution, while HRNet may be useful in limited-data settings or for features requiring high spatial resolution spatial preservation.