A deep learning architecture and workflow for geomorphic feature extraction using digital terrain model‐derived land surface parameters

Abstract An architecture and workflow are proposed and described for extracting geomorphic features from digital terrain model (DTM)‐derived land surface parameters (LSPs) using deep learning (DL)‐based semantic segmentation and integration of LSP calculations into the model architecture to allow for tensor‐ and graphics processing unit (GPU)‐based computation within the DL framework. To characterize terrain patterns at multiple spatial scales, the input DTM can be generalized using Gaussian pyramids (GPs). The LSPs are then provided to the trainable component of the model. The workflow is explored using two examples: valley fill faces resulting from mountaintop removal surface coal mine reclamation, an example of anthropogenic geomorphic features, and sinkholes within a karst landscape, an example of natural geomorphic features. We compare three different, ‘U’‐shaped architectures as the trainable component of the architecture: a traditional, convolutional neural network (CNN)‐based UNet, UNet with a ConvNeXt‐based encoder and attention gates along the skip connections and UNet with a Mamba‐based encoder and a CNN‐based decoder. As a baseline, DL models are compared with a pixel‐based random forest (RF) model using the same LSP feature space. We document improved performance in comparison with the pixel‐based approach and minimal differences among the three DL architectures based on F1 scores. The inclusion of GPs did not have a large impact on predictive performance for either the RF‐ or DL‐based models. Calculating LSPs as part of the model architecture did not greatly increase the computational complexity of the model. Depending on the model configuration, incorporating the LSP calculations into the model architecture increased the number of floating‐point operations (FLOPs) and multiply–accumulate operations (MACs) by 2.8% to 17.6% and increased the saved model size by 1.2 MB. In practice, only DTM data, mapping extent(s) and example features need to be provided during the training process. Disk space usage is decreased by a factor of 6 when not using GPs and 31 when incorporating these multiscale representations. The workflow offers an efficient means to extract geomorphic features from DTMs and can support downstream modelling and research tasks. It has been made available in the R language via the geodl package and Python via the terrainseg package.