State-Space-Based Mamba and Convolutional Neural Network Integration for Anthropogenic Geomorphic Feature Extraction from Land Surface Parameters

Accurate and scalable extraction of geomorphic features from LiDAR-derived terrain data is important for environmental monitoring, land use planning, and geohazard assessment. Recent deep learning advances have introduced state-space models, such as Mamba, which can capture broad spatial dependencies with linear complexity and offer an alternative to conventional convolutional neural network (CNN)- or Transformer-based approaches. This study evaluates Mamba-based architectures for geomorphic segmentation across three landform types: agricultural terraces, mine benches, and valley fill faces. Using high spatial resolution terrain derivatives, or land surface parameters (LSPs), as an input feature space, we compare a pure Mamba–UNet, a hybrid architecture with a Mamba-based encoder and CNN decoder, and a baseline CNN-based UNet with a ResNet-34 encoder, each trained with frozen and unfrozen encoders. Results indicate that no single architecture was best across all tasks; however, the hybrid Mamba-CNN architecture provided the most consistent performance overall, improving spatial continuity and interior feature recovery for tasks requiring broad spatial context, although its advantage over the baseline UNet was modest and task-dependent, and boundary-level analysis indicated that CNN-based configurations delineated feature edges most precisely, while the randomly initialized UNet remained highly competitive and achieved the strongest results for some tasks. Further, Mamba-based encoders converge quickly and maintain strong performance when frozen, whereas CNN-based UNets benefit more from encoder fine-tuning.