Automatic Landslide Detection Using Bi-Temporal Sentinel 2 Imagery

Abstract

Landslide inventory data sets are required for any landslide susceptibility mapping and prediction approaches. However, generating accurate landslide inventory data sets depends on applied methods and quality of input data, for example spatial resolution for satellite imagery. Therefore, the accuracy and availability of inventories vary in different studies. This study evaluated a strategy of sudden landslide identification product (SLIP) for landslide detection using Bi-Temporal Sentinel 2 Imagery and ALOS Digital Elevation Model (DEM). The resulting landslide detection map was then compared with an improved version of SLIP based on a fuzzy overlay. The resulting probability map was classified into three classes using the natural breaks method; the third class with the highest probability was extracted as the final map. The accuracy assessment stage demonstrated that using the improved version increased the accuracy by 16% compared to the SLIP method.

Detecting Coseismic Landslides in GEE Using Machine Learning Algorithms on Combined Optical and Radar Imagery

“…In this study, all these aspects were carefully investigated and addressed to achieve the best possible landslide detection accuracy in GEE while using ready-made datasets in GEE. Comparisons with existing approaches [19,29,[31][32][33][34][38][39][40][41][42][44][45][46][47][48] demonstrated superior performance of ML-LaDeCORsat of at least 10% higher detection accuracy.…”

“…(3) steep slopes, determined from DEM, to restrict the detection process to areas with pronounced topographic inclines; and (4) s land cover mask to minimize errors of commission specifically within recognized agricultural regions. SLIP was developed for MODIS and L8 imagery, adapted by integrating the inverse NDVI to assess the soil bareness (aSLIP) [33], and improved to utilize S2 instead of L8 imagery (iSLIP) [34]. In addition, Zhang, et al [35] discussed the potential presence of "old landslides"-areas of previously triggered landslides that did not (fully) recover at the time of the new incident.…”

“…Next, the spatial extent of sampling areas and the multi-band stacked input image were reduced by removing pixels within areas of low slopes, less than 10 degrees. Such slope filtering has been applied in many of the existing RS-based landslide detection approaches, including [29,[32][33][34]. This, however, can cause false positives (FP) for scenarios in which the landslide is triggered at steeper slopes, but its flow area reaches into areas of lower slopes, as illustrated in Figure 3.…”

The landslide traces inventory in the transition zone between the Qinghai-Tibet Plateau and the Loess Plateau: a case study of Jianzha County, China

“…With the development of automatic identification technology, it has a significant advantage in quickly obtaining regional landslide data. However, its accuracy may be not very good (Fayne et al, 2019;Zhang et al, 2020;Piralilou et al, 2021;Vecchiotti et al, 2021;Milledge et al, 2022). Combining the strengths of both approaches, the human-computer interaction visual interpretation of satellite images has gradually become an important method for constructing landslide inventory (Xu et al, 2015;Shao et al, 2020;Li et al, 2021;Cui et al, 2022a).…”

Detecting Co-Seismic Landslides in GEE using Machine-Learning Algorithms on combined Multispectral and RADAR Imagery

“…It is an multi-sensor approach that explores changes using four thresholds: 1) Increases in red wavelength band to signify the exposure of bare earth; 2) Variations in the Shortwave Infrared (SWIR) bands to indicate changes in soil moisture; 3) Steep slopes, determined from DEM, to restrict the detection process to areas with pronounced topographic inclines; and 4) A land-cover mask to minimize errors of commission specifically within recognized agricultural regions. SLIP had been developed for MODIS and L8 imagery, adapted by integrating the inverse NDVI to assess the soil bareness (aSLIP) [37] and improved to utilize S2 instead of L8 imagery (iSLIP) [38]. In addition, Zhang, et al [39] discussed the potential presence of "old landslides" -areas of previously triggered landslides that did not (fully) recover at the time of the new incident.…”