| Abstrakt: |
Landslides are complex geological phenomena influenced by various factors such as heavy rainfall, topography, geology, and anthropogenic activities. They can have devastating consequences, leading to prolonged road closures, costly detours, and, in severe cases, loss of human lives. For optimal management and planning of economic projects and construction, accurate and effective mapping of this hazard is essential. Deep learning has been successfully applied to landslide mapping, although this approach can occasionally face precision issues and requires significant time to determine the most suitable model for each data type. The originality of this research resides in the effectiveness and performance of the use of the neuronal architecture research (NAS) technique to optimize landslide susceptibility mapping. NAS proved to be the appropriate solution to these challenges through its ability to automate the architecture of deep learning models by optimizing its principal hyperparameters: the number of hidden layers (NCC), the number of neurons per hidden layer (NNC) and the number of training epochs (NEP). The Simulated Annealing meta-heuristic was employed to optimize the search space, enabling efficient exploration of various combinations in order to identify the best performing configuration for landslide modelling along the main roads in the province of Skikda, covering a 500-m-wide zone. To create a solid database, nine causal factors were selected by analyzing their relationship with landslide occurrence. These factors include slope, aspect, lithology, the Normalized Difference Vegetation Index (NDVI), soil type, elevation, as well as proximity to roads, watercourses, and geological faults. This selection was carried out using linear and non-linear statistical tests, based on Pearson's correlation coefficient and Spearman's ranks. The use of this technique shows an accuracy, recall, and F1-score of 0.9940, a precision of 0.9941, and an RMSE of 0.0772, which proves a better modeling of this phenomenon. The map produced by this approach indicates a good adequacy of the distribution of risk areas with the variability of susceptibility factors in the study area and the results obtained. The map displays different susceptibility classes: 22.1% of the area is categorized as very high susceptibility, 5.63% as moderate susceptibility, and 59.51% as very low susceptibility. Additionally, there are intermediate classes with 6.70% falling between very low and moderate susceptibility, and 6.06% between moderate and high susceptibility. The susceptibility map generated by this model, using the NAS method and the ArcGIS tool, provides vital information for identifying areas at high risk of landslides, which contribute to a better understanding of natural hazards and in the determining of the safety of residents and the preservation of road infrastructure in the Skikda province. |
Full text from SpringerLink