Landsifier v1.0: a Python library to estimate likely triggers of mapped landslides

• Published in: Natural hazards and earth system sciences Vol. 22; no. 11; pp. 3751 - 3764
• Main Authors: Rana, Kamal, Malik, Nishant, Ozturk, Ugur
• Format: Journal Article
• Language: English
• Published: Katlenburg-Lindau Copernicus GmbH 22.11.2022; Copernicus Publications
• Subjects: Accuracy; Algorithms; Archipelagoes; Artificial neural networks; Autocorrelation; Automation; Classification; Classifiers; Configurations; Data analysis; Data mining; Data processing; Deep learning; Digital Elevation Models; Earthquakes; Feature extraction; Geological hazards; Inventories; Landslides; Landslides & mudslides; Libraries; Machine learning; Methods; Model accuracy; Model testing; Modelling; Neural networks; Planforms; Polygons; Python; Testing; Topology; Training; Japan
• ISSN: 1684-9981; 1561-8633; 1684-9981
• DOI: 10.5194/nhess-22-3751-2022

Landslide hazard models aim at mitigating landslide impact by providing probabilistic forecasting, and the accuracy of these models hinges on landslide databases for model training and testing. Landslide databases at times lack information on the underlying triggering mechanism, making these inventories almost unusable in hazard models. We developed a Python-based unique library, Landsifier, that contains three different machine-Learning frameworks for assessing the likely triggering mechanisms of individual landslides or entire inventories based on landslide geometry. Two of these methods only use the 2D landslide planforms, and the third utilizes the 3D shape of landslides relying on an underlying digital elevation model (DEM). The base method extracts geometric properties of landslide polygons as a feature space for the shallow learner – random forest (RF). An alternative method relies on landslide planform images as an input for the deep learning algorithm – convolutional neural network (CNN). The last framework extracts topological properties of 3D landslides through topological data analysis (TDA) and then feeds these properties as a feature space to the random forest classifier. We tested all three interchangeable methods on several inventories with known triggers spread over the Japanese archipelago. To demonstrate the effectiveness of developed methods, we used two testing configurations. The first configuration merges all the available data for the k-fold cross-validation, whereas the second configuration excludes one inventory during the training phase to use as the sole testing inventory. Our geometric-feature-based method performs satisfactorily, with classification accuracies varying between 67 % and 92 %. We have introduced a more straightforward but data-intensive CNN alternative, as it inputs only landslide images without manual feature selection. CNN eases the scripting process without losing classification accuracy. Using topological features from 3D landslides (extracted through TDA) in the RF classifier improves classification accuracy by 12 % on average. TDA also requires less training data. However, the landscape autocorrelation could easily bias TDA-based classification. Finally, we implemented the three methods on an inventory without any triggering information to showcase a real-world application.