New NASA Landslide Mapping System Running at NCCS

Fig. 1: A section of the Pasang Lhamu Highway on the way to Dhunche, Langtang National Park, Rasuwa, Nepal. NASA and university scientists are studying hundreds of landslides along this highway in the Himalayas. Photo by Pnepalensis; acquired from Wikimedia Commons.

Empowered by NASA Center for Climate Simulation (NCCS) high-performance computing resources, NASA Goddard Space Flight Center and University of Twente scientists have demonstrated a new open-source landslide mapping system that is significantly faster than manual methods while achieving reasonable accuracy.

NASA’s Semi-Automatic Landslide Detection (SALaD) system combines three leading-edge technologies: open-source Python packages and modules, object-based image analysis (OBIA), and machine learning (ML). Satellite data and manual landslide maps serve as inputs for analysis.

The SALaD system runs on the NCCS ADAPT Science Cloud, using a Linux virtual machine with 10 computing cores and 50 gigabytes of memory.

SALaD’s first test was analyzing a 575-square-kilometer (km2) area along the Pasang Lhamu Highway in Nepal. This Himalayan region ranges from 604 to 4,341 meters (m), or 1,982 to 14,242 feet, in elevation. In 2015, the area had hundreds of landsides triggered by the magnitude-7.8 Gorkha earthquake, its magnitude-7.2 aftershock, and rainfall.

Fig. 2: The maps show the location of the study area along the Pasang Lhamu Highway in Nepal with manually mapped landslides. The yellow tile highlights the subset area used for machine learning training with NASA’s Semi-Automatic Landslide Detection (SALaD) system. Figure from Amatya et al. 2021.

To create landslide diagnostic features for the area, SALaD ingested 5-m resolution RapidEye satellite optical imagery and 30-m resolution NASADEM elevation data. NCCS ADAPT disk held 40 gigabytes of data from these platforms.

Focusing on 2015, the scientists also used the RapidEye imagery to manually map 623 landslides across the area. A 50-km2 zone (see the yellow tile in Fig. 2 above) with 148 landslides provided the data for ML-based training so SALaD could learn how to classify landslide and non-landslide objects. With a trained classifier in hand, SALaD used ML to classify the objects in the non-training zone.

Impact: An open-source landslide mapping system enables rapid identification of landslide locations after extreme events and helps guide transportation and distribution planning of resources to affected areas.

SALaD successfully detected 70% of the manually mapped landslide areas (see Fig. 3 below). On ADAPT, SALaD took 1 hour to generate a landslide map compared to 6 hours of manual mapping.

Fig. 3: These insets show landslides a. mapped manually and b. detected by the SALaD system. SALaD has overall accuracy of 70% compared to manual methods. It did well at detecting large landslides and landslide scarps (steep surfaces at the upper edges) but failed to detect long narrow runouts, which lack distinct diagnostic characteristics. Figure from Amatya et al. 2021.

“NASA supercomputing offers the computing power needed to rapidly map landslides following an extreme event,” said Pukar Amatya, associate scientist for Earth sciences with USRA and NASA Goddard’s Hydrological Sciences Laboratory. “Parallel processing capabilities to process images simultaneously across multiple virtual machines help create landslide maps across large areas within hours of images being available.”

Amatya and his colleagues are exploring SALaD improvements such as running the system on ADAPT’s Graphical Processing Units (GPUs) for even greater speed and using an ensemble of ML algorithms for better accuracy. Ongoing SALaD advances are enabling swift landslide detection and information sharing with emergency response agencies in other areas of the globe, including Central America.

Related Link

Jarrett Cohen, NASA Goddard Space Flight Center