Assessing Deep Learning Segmentation Models for Mapping Landslides From LiDAR‐Based Elevation Data
- 1Earth and Environmental Sciences University of Kentucky Lexington KY USA
- 2Kentucky Geological Survey University of Kentucky Lexington KY USA
- 3Computer Science and Engineering McKelvey School of Engineering Washington University in St. Louis St Louis MO USA
Abstract
Robust landslide inventories are critical for characterizing landslide hazard and helping reduce societal vulnerability. For this study, LiDAR‐derived digital elevation models (DEMs), their derivatives, and a landslide inventory were used to train different deep learning segmentation models to assess the functionality of automatic segmentation for high‐resolution, non‐event specific landslide inventories. Three deep learning segmentation approaches (U‐Net, MA‐Net, and FPN) were trained with a variety of hyperparameters, with some consideration of input modalities. Models were assessed with a spatial test holdout and then further evaluated against a proximal holdout landslide inventory. In total, 90 model scenarios were trained. Model performance metrics and visual assessment suggested that in general, model performances varied only slightly during hyperparameter tuning. U‐Net trained on DEM gradients performed the best, with a feature IoU (intersection over union) of 0.407 among all scenarios. MA‐Net trained on DEM gradients had the best performance against the proximal holdout inventory, with an IoU of 0.398, however these performances are within random variability of the models indicating no clear best model. A comparison of the proximal holdout inference results from various trained models suggested that deep learning segmentation models are effective in mapping landslides from LiDAR‐derived DEMs. A well‐trained landslide segmentation model could potentially be used to aid manual mapping by improving efficiency and accuracy. The direct inference result could also be used as an approximate inventory for use in data‐driven models, such as landslide susceptibility models, to significantly reduce the time and resources required for traditional landslide mapping.
Versions
Related
- Assessing the seasonal compartmentalization of water fluxes in the soil-plant-atmosphere continuum of a high-elevation mountain grassland
- A machine learning model for early season wildfire risk in Ghana’s Guinea savannah assessing discrimination, operational prevalence, spatial transferability, and CMIP6 climate scenarios
- A Minimal Model For Long Runout Landslides
- A Sentinel-2 benchmark dataset for deep-learning active-fire segmentation across 25 California wildfires, July 2020-August 2026
- Investigations on the initiation mechanism of rainfall-induced loess landslides driven by sinkhole-assisted seepage: a case study in Qinghai, China
BibTeX
@article{koch2026assessing,
title = {{Assessing Deep Learning Segmentation Models for Mapping Landslides From LiDAR‐Based Elevation Data}},
author = {Hudson J. Koch and Junfeng Zhu and Shashank Bhatt and Nathan Jacobs and Jason M. Dortch},
journal = {Earth and Space Science},
year = {2026},
month = oct,
doi = {10.1029/2025ea004937},
url = {https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025EA004937},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Assessing Deep Learning Segmentation Models for Mapping Landslides From LiDAR‐Based Elevation Data
AU - Hudson J. Koch
AU - Junfeng Zhu
AU - Shashank Bhatt
AU - Nathan Jacobs
AU - Jason M. Dortch
PY - 2026
DA - 2026/10/03
JO - Earth and Space Science
DO - 10.1029/2025ea004937
UR - https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025EA004937
AB - Robust landslide inventories are critical for characterizing landslide hazard and helping reduce societal vulnerability. For this study, LiDAR‐derived digital elevation models (DEMs), their derivatives, and a landslide inventory were used to train different deep learning segmentation models to assess the functionality of automatic segmentation for high‐resolution, non‐event specific landslide inventories. Three deep learning segmentation approaches (U‐Net, MA‐Net, and FPN) were trained with a variety of hyperparameters, with some consideration of input modalities. Models were assessed with a spatial test holdout and then further evaluated against a proximal holdout landslide inventory. In total, 90 model scenarios were trained. Model performance metrics and visual assessment suggested that in general, model performances varied only slightly during hyperparameter tuning. U‐Net trained on DEM gradients performed the best, with a feature IoU (intersection over union) of 0.407 among all scenarios. MA‐Net trained on DEM gradients had the best performance against the proximal holdout inventory, with an IoU of 0.398, however these performances are within random variability of the models indicating no clear best model. A comparison of the proximal holdout inference results from various trained models suggested that deep learning segmentation models are effective in mapping landslides from LiDAR‐derived DEMs. A well‐trained landslide segmentation model could potentially be used to aid manual mapping by improving efficiency and accuracy. The direct inference result could also be used as an approximate inventory for use in data‐driven models, such as landslide susceptibility models, to significantly reduce the time and resources required for traditional landslide mapping.
SN - 2333-5084
SN - 2333-5084
ER -