A CNN-based mapping approach for aeolian dunes in Arctic Sweden

Aeolian dunes provide valuable evidence of past wind regimes and environmental change but mapping them reliably over large areas remains challenging. Manual delineation of dune boundaries is time-consuming and difficult to apply consistently. Automated approaches also struggle where vegetation partly obscures dune morphology and where sand-rich fluvial surfaces, such as riverbanks and terraces, exhibit spectral responses similar to those of stabilized dunes. In this study, dune mapping is treated as a pixel-wise binary segmentation task that separates dune-covered areas, including adjoining or overlapping dunes, from the surrounding background. We present ArcticSandNet, a multimodal segmentation network based on the U-Net architecture. Optical imagery and terrain variables derived from a digital elevation model (DEM), including elevation, slope, and curvature, are processed by two separate encoders. Their features are combined at each encoder level through feature concatenation. In the decoder, an attention-based module fuses encoder skips features with upsampled decoder features, helping the network retain dune-related spatial information while suppressing responses from heterogeneous background surfaces. The model was trained and evaluated across 15 sites in northern Sweden using spatial five-fold cross-validation. A further site was excluded from all model training, validation, and selection procedures and retained solely for independent testing. In cross-validation, ArcticSandNet achieved a two-class mean intersection over union (mIoU) of 78.15% ± 5.34% and a two-class mean F1 score (mF1) of 85.97% ± 4.78%. For the dune class, it achieved an F1 score of 72.61% ± 8.46% and an IoU of 57.63% ± 9.59%, outperforming U-Net, SegNet, and Attention U-Net trained under the same settings. On the independent test site, ArcticSandNet achieved a dune-class recall of 60.52%, an F1 score of 67.16%, and an IoU of 50.55%, also outperforming the three benchmark networks in dune-class F1 and IoU. The resulting maps of dune-covered areas provide a consistent spatial basis for regional analyses of dune-field characteristics, morphology, and landscape stability.