PREPARING HIGH-QUALITY DEEP LEARNING TRAINING DATASETS FOR LANDSLIDE MAPPING IN WESTERN NORTH CAROLINA

The North Carolina Geological Survey is in the early-stages of training deep learning (DL) models to detect and map landslides from a lidar DEM. During the initial preparation of landslide-mapping training and validation datasets for DL applications, we identified several areas for improving data quality in the existing landslide inventory.

Preliminary data interrogation and exploratory data analysis revealed data quality concerns that we are working to ameliorate. We identified four key areas for improvement: clarifying definition of terms, enhancing the mapping protocol, filling data gaps, and assessing mapping accuracy.

Some of these data quality issues are manageable with data cleaning and wrangling while others are less manageable. Issues related to the mapping protocol are particularly challenging, highlighting the value of a well-documented protocol (i.e., “DOGAMI Protocol”) in the development of landslide inventories. Issues that cannot be resolved with cleaning and wrangling will either be excluded from DL training or will be re-mapped. We anticipate that trained DL models will improve the precision, accuracy, and mapping efficiency of our landslide inventory over the coming decade.