Semi-automatic mapping of shallow landslides using free Sentinel-2 and Google Earth Engine

Abstract

Abstract. The global availability of Sentinel-2 data and the widespread coverage of free-cost and high-resolution images nowadays give opportunities to map, at low-cost, shallow landslides triggered by extreme events (e.g. rainfall, earthquake). A rapid and low-cost shallow landslides mapping could improve damages estimations, susceptibility models or land management. This work presents a semi-automatic methodology to map potential landslides (PL) using Sentinel-2 images, and it is the first step toward more detailed mapping. We create a GIS-based and user-friendly methodology to extract PL based on pre- post- event NDVI variation and geomorphological filtering. The semi-automatic inventory was compared with benchmark landslides inventory drawn on high-resolution images. We also used the Google Earth Engine scripts to extract the NDVI time series and make a multi-temporal analysis. We apply this to two study areas in NW Italy hatted in 2016 and 2019 by extreme rainfall events. The results show that the semi-automatic mapping based on Sentinel-2 allows detecting the majority of shallow landslides larger than satellite ground pixel (100 m2). PL density and distribution well match with the benchmark. However, the false positives (30 % to 50 % of cases) are challenging to filter, especially when they correspond to river bank erosions or cultivated land.

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

“…Most methods include pixel-based slope data in their algorithms and several approaches mask pixels below a defined slope threshold, excluding them from potential landslide pixels. In [33] for instance, a slope threshold of 15 degrees was applied for masking. Some methods also consider Principal component analysis (PCA) that transforms multiple spectral bands acquired at different times into distinct linear components which results in noise reduction and removal of redundant information in the data, ultimately enhancing the distinction between spectral bands [34,35].…”

“…A rise in published journal articles related to GEE over the last three years highlights the increased popularity of GEE, with L8 and S2 being the most widely used earth observation satellite sensors, and articles based on Random Forest and Water Resources most often reported [58]. Several existing landslide detection methods have been implemented in GEE [14,23,33,[45][46][47][48][49][50]59]. However, only Handwerger et al [49] provide a shared GEE script that allows replication and potential adjustment of the script to other landslide events.…”

“…Next, the spatial extent of sampling areas and the multi-band stacked input image were reduced removing pixels within areas of low slopes, less than 10 degrees. Such slope-filtering is being applied in many of the existing RS-based landslide detection approaches, including [33,[36][37][38]. 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.…”

“…Utilizing P2 polarization bands, the same process as for S1 was applied and the following band was identified to be included in the ML classifier used in this study: P2_log_HV, the log ratio for preand post-event P2 SAR (HV) backscatter. Derived from GDEM, the following topographic factors were included: slope, aspect, curvature, and elevation [33,35,42]. These topographic data are derived from imagery taken prior to the earthquake event.…”

Rapid Mapping of Landslides Induced by Heavy Rainfall in the Emilia-Romagna (Italy) Region in May 2023

“…Some operative choices represent a trade-off between the rapidity of execution and data quality/resolution, for instance, the mapping was realized at a 1:5000 scale and should be consulted at most at the same scale; a side effect is that our mapping is not recommended for higher resolution susceptibility studies, unless a field verification is undertaken. We underline that our inventory is one of the few that have been publicly released; to our knowledge only one other inventory has been published to date, realized through automatic mapping from Sentinel-1 images ( [50]; publication date 1 June 2023).…”