Application of deep learning method in automatically detecting rainfall-induced shallow landslides in a data-sparse context

Abstract

Detecting rainfall-induced shallow landslides in data-sparse contexts has become an environmental concern in recent decades and is crucial for a comprehensive landslide disaster management plan (CLDMP). Most of the previous works have contributed to the development of automated methods for detecting earthquake-triggered landslides. Despite the substantial contributions of researchers in this field, gaps and uncertainties still exist in developing a method for automatically detecting rainfall-induced shallow landslides. To address this gap, the present study has utilized the deep learning (DL) based U-net model for automatically detecting rainfall-induced shallow landslides from multi-temporal, very high-resolution (VHR) PlanetScope, medium resolution (MR) Sentinel-2 imagery, and ALOS PALSAR-provided digital elevation model (DEM), collected from the years 2018, 2019, 2022, and 2023. Four different data sets have been prepared for this study: Dataset A, comprising red, green, blue (RGB), and near-infrared (NIR) bands of PlanetScope imagery; Dataset B, comprising RGB and NIR bands of PlanetScope imagery with the inclusion of the normalized difference vegetation index (NDVI) calculated from the red and NIR bands, elevation, and slope derived from DEM; Dataset C, comprising RGB and NIR bands of Sentinel-2 imagery; and Dataset D, comprising RGB and NIR bands of Sentinel-2 imagery with the inclusion of NDVI, elevation, and slope. As a case study, the Chittagong Hill Tracts (CHT) of Bangladesh have been selected. For training the U-net model with ground truth data, 181 landslide polygons have been created from Google Earth Pro, which is a small set of ground truth data. So, the horizontal flip technique has been applied to augment the dataset, effectively doubling the entire dataset. Each dataset (A, B, C, and D) has been experimented with in 4 different trials utilizing the repeated stratified hold-out validation method so that all data is used as test data, to avoid biased results. Comparatively, Trials 1 and 2 contain a larger set of landslide training samples than Trials 3 and 4. Thus, 16 different experiments have been conducted in the present study. The performance of the U-net model is evaluated by precision, recall, F1 score, loss, and accuracy metrics. It is explored from the experiment that Datasets A and B perform the best; however, the integration of the DEM data does not enhance the accuracy of the model. The datasets comprised of Sentinel-2 imagery (Datasets C and D) exhibited very poor performance in all trials (4) in detecting rainfall-induced shallow landslides. Among the four Trials, utilizing Dataset A and B, Trials 1 and 2 outperformed, indicating the necessity of using larger training samples for DL model implementation. The mean precision, recall, F1 score, loss, and accuracy based on Trials 1 and 2 are 1, 0.625, 0.625, 0.380, and 0.999, respectively (same results found in both Datasets A and B). Overall, the performance of the model indicates that the U-net model can be used to detect rainfall-induced shallow landslides across similar geographic regions and temporal contexts around the world