Comprehensive performance assessment of landslide susceptibility mapping with MLP and random forest: a case study after Elazig earthquake (24 Jan 2020, Mw 6.8), Turkey


ENVIRONMENTAL EARTH SCIENCES, vol.81, no.5, 2022 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 81 Issue: 5
  • Publication Date: 2022
  • Doi Number: 10.1007/s12665-022-10225-y
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, IBZ Online, PASCAL, Aerospace Database, Agricultural & Environmental Science Database, Applied Science & Technology Source, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), BIOSIS, CAB Abstracts, Compendex, Computer & Applied Sciences, Environment Index, Geobase, INSPEC, Pollution Abstracts, Veterinary Science Database, Civil Engineering Abstracts
  • Keywords: Landslide susceptibility mapping, Performance assessment, Machine learning, Photogrammetry, Multi-layer perceptron, Random forest, SPATIAL-DISTRIBUTION, LOGISTIC-REGRESSION, REGION, PREDICTION, MODELS, MAPS, TREE
  • Hacettepe University Affiliated: Yes


Quality assessment (QA) for landslide susceptibility maps (LSMs) is essential to increase their usability. A QA approach based on the landslide activity after a triggering event can be useful for the performance evaluation of the methods used for LSM production. Landslides triggered by earthquakes can be employed for this purpose as they occur frequently throughout the active seismic regions of the world. After an earthquake occurred in Elazig, Turkey on 24 Jan 2020 (Mw 6.8), several landslides were activated in the mountainous parts. Here, the performances of two state-of-the-art machine learning methods, i.e., the random forest (RF) and the multi-layer perceptron (MLP), were investigated using the activated landslides. The landslide inventory was derived in a previous study by using pre- and post-event aerial photogrammetric datasets and classified according to their activity types and temporal observations. The classes observed in the pre-event photogrammetric datasets were inactive (L1) and active mass movements (L2). The ones observed in the post-event photogrammetric datasets were new active zones inside the existing landslide (L3) and new activity (L4). Here, only the L1 and L2 type landslides observed in a part of the study area were used for the model training and the LSMs were produced for the whole area to investigate the model transferability. The L3 and L4 type landslides were used for validation. In addition, the area under curve (AUC) values obtained from the methods and the volumetric change maps obtained from the pre- and post-event digital elevation models were also used for the performance assessment. The results demonstrated that RF exhibited higher classification accuracy (AUC = 0.93) than MLP (AUC = 0.87); and accurate LSMs could be produced by using a sub-part of the basin for training.