Slide type landslide susceptibility assessment of the Buyuk Menderes watershed using artificial neural network method


Tekin S., ÇAN T.

ENVIRONMENTAL SCIENCE AND POLLUTION RESEARCH, 2022 (Peer-Reviewed Journal) identifier identifier identifier

  • Publication Type: Article / Article
  • Publication Date: 2022
  • Doi Number: 10.1007/s11356-022-19248-1
  • Journal Name: ENVIRONMENTAL SCIENCE AND POLLUTION RESEARCH
  • Journal Indexes: Science Citation Index Expanded, Scopus, IBZ Online, ABI/INFORM, Aerospace Database, Agricultural & Environmental Science Database, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), BIOSIS, CAB Abstracts, EMBASE, Environment Index, Geobase, MEDLINE, Pollution Abstracts, Veterinary Science Database, Civil Engineering Abstracts
  • Keywords: Buyuk Menderes watershed, Western Anatolia, Landslide susceptibility, Artificial neural network, GIS, TURKEY, COUNTY, SCALE

Abstract

The Buyuk Menderes watershed is the largest drainage watershed in Western Anatolia with an area of approximately 26,000 km(2). In the study area, almost 863 landslides occurred, extending over 222 km(2) with a mean landslide area of 0.21 km(2). In this study, landslide susceptibility assessments were carried out using artificial neural network method, which is one of the data-driven methods. In this study, that will contribute to the mitigation or control of the landslides caused by the reasons controlling the spatial and temporal distribution of landslides created in the GIS and MATLAB environment by using scientific and technological approaches within the framework. Since derivative activation function is also used in back-propagation artificial neural networks, its derivative is easily calculated in order not to slow down the calculation. Levenberg-Marquardt back-propagation (LM), resilient back propagation back-propagation (trainrp), scaled conjugate gradient back-propagation (trainscg), conjugate gradient with Powell/Beale restarts back-propagation (traincgb), and Fletcher-Powell conjugate gradient back-propagation (traincgf) algorithms are used, which constantly interrogate the link between the input parameter and the result output, and at least one cell's output is given as an input to any other cell. Geology, digital elevation model, slope, topographic wetness index, roughness index, plan, profile curvatures, and proximity to active faults and rivers were used as landslide conditioning factors. In susceptibility assessments, landslides were separated by 70% analysis, 15% test, and 15% validation datasets by random selection method. The performances of the landslide susceptibility maps were assessed by the area under the ROC curve (AUC), accuracy (ACC), precision, recall, F1 score, Kappa test error histogram, and confusion matrix, respectively. The area under the receiver operating characteristic curves, analysis, testing, validation, landslides, and study areas were found between 0.873 and 0.911. The susceptibility map had a high prediction rate in which high and very high susceptible zones corresponded to 26% of the study area including 82% of the recorded landslides.