Event landslide mapping using L-band SAR data: insights from the 2018 Hokkaido earthquake

Event landslide mapping plays a critical role in understanding the impact of triggering events, supporting emergency response, and defining risk reduction strategies. It also provides valuable datasets for validating susceptibility and risk models and training automatic landslide detection methods using machine learning. Enhancing our ability to detect and map landslides, particularly under challenging conditions, is key to improving response capacity during large-scale events.

Optical post-event images, while commonly used for mapping, are often unavailable immediately after a disaster due to dense cloud cover and limited revisit times. Synthetic Aperture Radar (SAR) sensors, with their ability to acquire data regardless of cloud cover or lighting conditions, offer a promising alternative.

In this study, we evaluated the reliability of L-band ALOS SAR amplitude images to prepare a landslide inventory map for the region affected by the MW 6.6 Hokkaido earthquake in September 2018, which triggered over 6,000 landslides. Using amplitude images of the radar backscattering coefficient (beta naught), we derived log-ratio change detection maps that highlight surface changes caused by landslides. These maps were visually interpreted by an expert geomorphologist to produce a SAR-based inventory in three test areas selected to represent varying landslide densities: low, medium, and high, as defined by the benchmark inventory.

To validate this approach, we compared the SAR-based inventory with a benchmark inventory derived from the interpretation of post-event optical images and field checks. The comparison assessed spatial coverage, geometric accuracy, completeness, and size distribution.  

Results showed a good agreement between the SAR-based inventory and the benchmark, largely due to the high resolution of ALOS images, which enabled accurate detection and delineation of most landslide-affected areas. However, in the high-density test area, the delineation of individual landslides was less precise, with some generalizations observed. In contrast, the low-density test area exhibited more commission errors, likely due to challenges in distinguishing true landslides from noise in sparsely affected regions.

Our findings further demonstrate the potential of SAR data for landslide mapping in complex scenarios. The robust dataset produced in this study provides a rational basis for developing and training automatic landslide mapping systems based on radar backscattering log-ratio images.