Development and Validation of the Japan High-Resolution Landslide Dataset (JHRLD): Integrating Remote Sensing and Deep Learning

Landslides are geomorphological hazards triggered by natural factors such as rainfall, seismic activity, and snowmelt. A landslide inventory is essential for understanding and assessing the processes, distribution, and risks associated with these events. While traditional manual mapping from aerial imagery delivers high accuracy for slope-scale studies, it is labor-intensive and impractical for large-scale applications. Advances in remote sensing technologies, including high-resolution satellite imagery and synthetic aperture radar (SAR), have significantly enhanced the efficiency of landslide detection over broader geographical scales. Concurrently, deep learning techniques, such as convolutional neural networks (CNNs), have revolutionized the field by automating landslide feature extraction and segmentation through remote sensing imagery, addressing the inefficiencies of manual methods.

Current challenges include the regional distribution bias in remote sensing-based landslide datasets, with limited high-quality data available for regions like Japan. Additionally, the lack of systematic evaluation of optimal features and deep learning architectures hinders improvements in detection accuracy and model transferability.

To address these gaps, we developed and validated the Japan High-Resolution Landslide Dataset (JHRLD), which integrates multi-sensor data encompassing spectral, SAR, and topographic features. The dataset comprises two subsets: Sentinel-2 for moderate-resolution (10 m) and PlanetScope for high-resolution (3 m) imagery, named after the optical images used for landslide delineation. Both subsets were designed based on a pool of 21 candidate features, including spectral bands, vegetation indices, SAR-derived backscatter metrics from Sentinel-1, and topographic attributes derived from the DEM published by the Geospatial Information Authority of Japan. A rigorous feature selection includes statistical and model-based evaluations, narrowing the list to the most significant features for landslide mapping, including green, red, NDVI, slope, and intensity.

Three deep learning models were employed on the JHRLD: UNet++, DeepLabv3+, and Medical Transformer (MedT). These models were evaluated using the F1 score for evaluating the JHRLD’s robustness and reliability. Performance analysis revealed that each model exhibited unique strengths depending on dataset resolution. On the moderate-resolution Sentinel-2 dataset, UNet++ excelled in detecting smaller-scale landslides, achieving an F1 score of 0.70. In contrast, DeepLabv3+ performed best on the high-resolution PlanetScope dataset, achieving an F1 score of 0.69 and effectively capturing large-scale and complex features. MedT showcased its superiority in boundary delineation, achieving the F1 score of 0.70 and excelling in identifying intricate landslide features. These results affirm the JHRLD’s robustness and reliability, providing a strong foundation for high-precision landslide detection across diverse resolutions and environments.

The JHRLD was validated in the Noto Peninsula, a region impacted by an Mw 7.5 earthquake and torrential rainfall in 2024. The model trained on the JHRLD demonstrated strong transferability, achieving an F1 score of 0.65 and a detection rate of 81% on this unseen area. Temporal and spatial analyses confirmed the JHRLD’s robustness, aligning well with observed hazard patterns under varying triggers.

The findings highlight the JHRLD’s adaptability as a benchmark dataset and its social utility in disaster prevention planning and emergency response.

 

Figure. Workflow for the JHRLD development and validation.