Search

Many SAR images have been utilized for geologic disasters investigations with the continuous launch of new Synthetic Aperture Radar (SAR) satellites such as ALOS-2/PALSAR-2. However, to proactively respond to transient slope failures caused by heavy rainfall, rapid extraction of areas of surface change accompanying slope failures is required. This study proposes two methods for quantitatively extracting slope failure areas using L-band SAR observations with slope units (SUs) as the evaluation units. The first method is based on the threshold method, which automates the selection of thresholds for various disaster-affected conditions, such as land use and topography. The second method is a machine-learning-based density ratio estimation method, which uses multi-temporal periodic observation data and pre- and post-disaster data to detect outliers through feature selection optimization. In the observation direction with the shortest satellite observation period, the F1 score (The F1 score is the harmonic mean of the precision and recall) of the threshold method for accuracy evaluation is 61.91%, and the F1 score of the density ratio method is 65.87%. Both methods can reduce the problem of low extraction accuracy caused by the effect of speckle noise. When slope failure occurs, both methods can extract the area of surface change within hours of a disaster. The method proposed in this study displays good applicability in supporting emergency rescue and the prevention of secondary disasters. [ABSTRACT FROM AUTHOR]

Abstract Many SAR images have been utilized for geologic disasters investigations with the continuous launch of new Synthetic Aperture Radar (SAR) satellites such as ALOS-2/PALSAR-2. However, to proactively respond to transient slope failures caused by heavy rainfall, rapid extraction of areas of surface change accompanying slope failures is required. This study proposes two methods for quantitatively extracting slope failure areas using L-band SAR observations with slope units (SUs) as the evaluation units. The first method is based on the threshold method, which automates the selection of thresholds for various disaster-affected conditions, such as land use and topography. The second method is a machine-learning-based density ratio estimation method, which uses multi-temporal periodic observation data and pre- and post-disaster data to detect outliers through feature selection optimization. In the observation direction with the shortest satellite observation period, the F1 score (The F1 score is the harmonic mean of the precision and recall) of the threshold method for accuracy evaluation is 61.91%, and the F1 score of the density ratio method is 65.87%. Both methods can reduce the problem of low extraction accuracy caused by the effect of speckle noise. When slope failure occurs, both methods can extract the area of surface change within hours of a disaster. The method proposed in this study displays good applicability in supporting emergency rescue and the prevention of secondary disasters.