FEATURE SPACE CONSIDERATIONS FOR GEOMORPHIC DEEP LEARNING USING DIGITAL TERRAIN VARIABLES

The impact of input feature space for extracting geomorphic features from raster-based digital terrain variables using convolutional neural network (CNN)-based semantic segmentation deep learning was explored for multiple use cases: mapping of valley fill faces resulting from mountaintop removal coal mining, agricultural terraces for erosion control, and alluvial deposits and thick glacial till surficial geology. We compared a three-band composite consisting of topographic position indices and the square root of slope to single-band hillshade, multidirectional hillshade, and slopeshade rasters. The three-band composite raster generally provided lower overall losses for the training and validation samples across training epochs and better performance for generalizing to withheld testing data. Our results suggest that CNN-based DL for mapping geomorphic features or landforms from digital terrain variables is sensitive to input feature space; as a result, researchers and analysts should consider the feature spaced provided to DL algorithms. We suggest that the three-band composite implemented here can offer better performance in comparison to using hillshades or other common terrain visualization surfaces.