Evaluation of machine learning-based algorithms for landslide detection across satellite sensors for the 2019 Cyclone Idai event, Chimanimani District, Zimbabwe

Das, Raja,Wegmann, Karl W.
DOI: https://doi.org/10.1007/s10346-022-01912-9
IF: 6.153
2022-08-13
Landslides
Abstract:Changes in climatic patterns, manifested as intensified cyclones and torrential rainfalls in a warming world will inevitably impact the frequency of landslides. One such climatic extreme, Cyclone Idai (March 2019), caused significant havoc across southeastern Africa, including Mozambique, Malawi, and eastern Zimbabwe, by triggering thousands of landslides and widespread concurrent flooding, both of which resulted in substantial loss of life. The study was conducted in the Chimanimani District of eastern Zimbabwe to understand the impact of Cyclone Idai on landslide initiation, quantify the volume of mobilized hillslope material during the event, and compare the event-triggered material release against the annual erosional yield across the study area using machine learning and remote sensing techniques. We evaluated satellite imagery of various resolutions, namely, PlanetScope (3m/px), RapidEye (5m/px), and Sentinel-2 (10m/px) and three machine learning algorithms: artificial neural network (ANN), random forest (RF), and support vector machine (SVM) to identify landslides and compared the efficacy of the models. A total of nine predictor variables derived from the satellite imagery and a 30-m ASTER Global Digital Elevation Model were employed to identify landslides and differentiate them from concurrent hydrological flooding. The models classified the study area into three classes: (i) landslides, (ii) flooding, and (iii) unaffected area. The RF model using PlanetScope satellite data attained the highest prediction accuracy of 97.88%, whereas the accuracy of other machine learning model-satellite data combinations ranged between 94.58 and 97.27%. Subsequently, landslide size thresholds were applied on the initially mapped landslides to eliminate noise and uncertainty from the data before estimating the final Cyclone Idai event-triggered landslide volume. A probability density function, which corresponds to a logarithmic plot of non-cumulative landslide frequency against the mapped landslide area, was employed to calculate landslide size thresholds using divergence and rollover point cutoff values. Finally, a landslide area-to-volume power-law scaling relationship was exploited to derive landslide volumes in the study area that ranged between $$6.8 \times 10^{6}$$ to $$14.7 \times 10^{6}$$ m3 and $$9.0 \times 10^{6}$$ to $$19.6 \times 10^6$$ m $$^3$$ for divergence and rollover point thresholds, respectively, across the different combinations of machine learning models and satellite sensors. The estimated landslide volume indicates hillslope material liberated by Cyclone Idai was 269 to 345 times greater than the estimated annual average background denudation of the study area computed from a topography-based local erosion model.
geosciences, multidisciplinary,engineering, geological
What problem does this paper attempt to address?