The SWADE model for landslide dating in time series of optical satellite imagery

Abstract

Landslides are destructive natural hazards that cause substantial loss of life and impact on natural and built environments. Landslide frequencies are important inputs for hazard assessments. However, dating landslides in remote areas is often challenging. We propose a novel landslide dating technique based on Segmented WAvelet-DEnoising and stepwise linear fitting (SWADE), using the Landsat archive (1985–2017). SWADE employs the principle that vegetation is often removed by landsliding in vegetated areas, causing a temporal decrease in normalized difference vegetation index (NDVI). The applicability of SWADE and two previously published methods for landslide dating, harmonic modelling and LandTrendr, are evaluated using 66 known landslides in the Buckinghorse River area, northeastern British Columbia, Canada. SWADE identifies sudden changes of NDVI values in the time series and this may result in one or more probable landslide occurrence dates. The most-probable date range identified by SWADE detects 52% of the landslides within a maximum error of 1 year, and 62% of the landslides within a maximum error of 2 years. Comparatively, these numbers increase to 68% and 80% when including the two most-probable landslide date ranges, respectively. Harmonic modelling detects 79% of the landslides with a maximum error of 1 year, and 82% of the landslides with a maximum error of 2 years, but requires expert judgement and a well-developed seasonal vegetation cycle in contrast to SWADE. LandTrendr, originally developed for mapping deforestation, only detects 42% of landslides within a maximum error of 2 years. SWADE provides a promising fully automatic method for landslide dating, which can contribute to constructing landslide frequency-magnitude distributions in remote areas.