Multi-Band Geomorphic Feature Design for Landslide Detection Using Mask R-CNN

Multi-Band Geomorphic Feature Design for Landslide Detection Using Mask R-CNN
Williams, J.M.1*, Winther, M.1,2, and Svennevig, K.1
*[email protected]
1Geological Survey of Denmark and Greenland (GEUS), Copenhagen, Denmark
2Department of Geosciences and Natural Resource Management, University of Copenhagen, Copenhagen, Denmark
Manual mapping of landslides in low-relief, post-glacial terrains such as Denmark remains subjective, time-intensive, and difficult to reproduce at national scale. The first national landslide inventory of Denmark, derived from a 0.4 m resolution digital elevation model and high-resolution orthophotos, identified over 3200 landslides through expert-based interpretation with an estimated completeness of ~87% (Luetzenburg et al., 2021; Svennevig et al., 2024). While this dataset provides a critical baseline for hazard assessment and geomorphic analysis, its manual construction highlights the need for scalable, reproducible, and updateable mapping approaches, motivating the application of machine learning methods.
Here, we evaluate how multi-band, physically derived terrain representations influence landslide instance segmentation using a Mask R-CNN framework. We construct a multi-band geomorphic feature stack derived from the national digital elevation model, integrating eight-directional hillshade composites with slope (degrees), multi-scale topographic position indices (50 m TPI and 300 m TPI), topographic wetness index (TWI), and profile curvature, with the digital elevation model optionally included.
This representation captures complementary morphometric information across spatial scales and terrain orientations, enabling improved detection of subtle landslide signatures in low-relief environments. Training data were derived from the Danish national landslide inventory, yielding 14,798 image chips (512 × 512 pixels) and 26,252 labeled landslide instances. A Mask R-CNN model with a ResNet-50 backbone was trained for 30 epochs within the ArcGIS Pro deep learning framework. Training and validation loss decreased consistently, and mean average precision exceeded 0.30 within early epochs, indicating effective feature utilization from the multi-band inputs.
To assess the contribution of individual input features, we implemented ablation experiments and permutation-based feature importance analysis, systematically removing or perturbing individual bands and quantifying changes in detection performance (He et al., 2016; Khallouk et al., 2025; Lundberg & Lee, 2017). Preliminary results suggest that multi-directional hillshades and multi-scale TPI provide the strongest contributions, while TWI and DEM improve detection stability in low-relief and hydrologically influenced settings.
The model successfully detects landslides across a range of morphologic expressions, including degraded and low-relief features underrepresented in existing inventories. These results demonstrate that input feature design is a critical control on deep learning performance in
geomorphic applications and provide a reproducible framework for improving national landslide inventories.
References:
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770–778. http://openaccess.thecvf.com/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html
Khallouk, N., Mabrouki, J., Tarik, L., El-Arrouch, T., & Alaoui, E. A. A. (2025). Explainable AI for gradient boosting-based relative humidity prediction in Fez, Morocco.
Luetzenburg, G., Svennevig, K., Bjørk, A. A., Keiding, M., & Kroon, A. (2021). A national landslide inventory of Denmark. Earth System Science Data Discussions, 2021, 1–13.
Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30. https://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html
Svennevig, K., Koch, J., Keiding, M., & Luetzenburg, G. (2024). Assessing the impact of climate change on landslides near Vejle, Denmark, using public data. Natural Hazards and Earth System Sciences, 24(6), 1897–1911.