SegFormer and SegFormer-UNet for anthropogenic geomorphic feature extraction from land surface parameters

Accurate, scalable mapping of anthropogenic geomorphic features from high spatial resolution terrain data remains challenging. Convolutional neural networks (CNNs) characterize local texture well but have limited receptive fields, whereas transformer-based architectures such as SegFormer model longer-range dependencies; their value for geomorphic mapping has not been systematically tested. We compare three architectures for pixel-level extraction of features from light detection and ranging (Lidar)-derived land surface parameters (LSPs): SegFormer, SegUNet (a hybrid SegFormer-UNet model), and a fully CNN-based UNet with a ResNet-34 encoder. Models are trained on 512 × 512 LSP chips for three tasks: agricultural terraces, historic mine benches, and valley fill faces. We evaluate random versus ImageNet initialization and, when pretraining is used, frozen versus unfrozen encoders, using focal Tversky loss to address class imbalance. Performance is assessed with F1-score, precision, recall, overall accuracy, intersection-over-union (IoU) and boundary IoU, with robustness summarized by resampling test chips; we also report input LSP importance and computational cost. End-to-end training is the most consistent driver of gains across backbones and datasets, particularly for SegFormer. UNet-style architectures with CNN-based decoders generalize most reliably, achieving the best precision and recall balance and the smoothest optimization. Contrary to common practice, random initialization often matches or exceeds ImageNet pretraining for LSP inputs, a result consistent with a mismatch in channel semantics: LSP bands are physically derived terrain measures rather than colour channels, so filters learned across RGB channels transfer less readily. Apparent difficulty tracks the proportion of boundary to interior pixels rather than geomorphic complexity: narrow agricultural terraces score lowest, mine benches are intermediate, and compact valley fill faces highest, although boundary IoU shows that valley fill edges are delineated least precisely. We therefore recommend unfrozen encoders with a CNN-based decoder as a strong default, and caution against assuming that ImageNet pretraining will help when using LSP predictor variables.