Uncertainty Reduction of Unlabeled Features in Landslide Inventory Using Machine Learning t-SNE Clustering and Data Mining Apriori Association Rule Algorithms

Abstract

A landslide inventory, after an intense rainfall event in 1998, Southwestern Korea, was collected by digitizing aerial photographs. This left high uncertainty in the inventoried features to be verified by ground truths. To reduce the uncertainty, the photographs were reexamined, supported by the time slider in Google Earth. We observed 77 deformed slopes, which were similar in shape and texture, to the inventoried landslides. We then sought to label the observed formations based on their spatial relationship with surrounding conditions. A three-phase methodology was developed. First, an inventory of landslide, no landslide, vulnerable slopes, and unlabeled features was analyzed based on spatial cluster patterns, and then the dimension was reduced using the t-distributed stochastic neighbor embedding (t-SNE). Second, the Apriori algorithm, based on association rule mining, was used to identify common relations in the inventory using landslide antecedent factors (derived from topographic and landcover maps) that are linked to areas of unlabeled features. Third, the findings were validated using Landsat TM (Thematic mapper) and ETM+(Enhanced thematic mapper) images acquired before and after the original inventory. Current research offers practical and economical solutions (reduced reliance on paid remote sensing sensors and field survey) to labeling and classification of missing or outdated spatial attributed information.