Comparative analysis and integration of knowledge- and data-driven multicriteria models: A case study of landslide susceptibility assessment in northern Iran
Abstract
A landslide susceptibility index (LSI) can be used to identify the susceptible areas to landslides and prevent loss of lives and infrastructure. LSIs are generated in knowledge- or data-driven landslide susceptibility assessment (LSA) models. Literature review shows that neither of these models is the best for LSA, and both should be integrated. Accordingly, some studies are being done in this regard, but with some drawbacks. This study's objective is to integrate and compare knowledge- or data-driven models within a standard LSA framework. For constructing the framework, the reference point (RP) method from Spatial Multicriteria Analyses (SMCA) set was employed. This method can combine geographic data and expert knowledge for hybridizing models, has a power parameter that allows for modifying LSIs, and benefits from a dual-susceptibility-measurement system. The framework was tested in a landslide-prone watershed in Iran called Fereydunkenar by determining the weights and value functions of causative criteria (e.g., slope degree) in knowledge-driven, data-driven, and intermediate forms. By combining the weighted value functions differently within the framework, nine types of RP-LSA models were created. Each model was repeated using three RP power parameters (q = 1, 2, ∞); 27 different LSIs were generated and compared by measuring their fitting certainty (FC) and prediction certainty (PC) using the receiver operating characteristics (ROC) method. The sensitivity of the models to criteria weights and value functions were also analyzed in global (all parameters at the same time) and local (one parameter at a time) forms by Monte Carlo simulation programmed in Python. Some of the created hybrid models were unreliable and highly sensitive, while others were successful. The hybrid model constructed from the knowledge-driven weights and average value functions was the best. The FC of this model was much better than that of the pure knowledge-driven model, and its PC was considerably higher than that of the pure data-driven model, although it was a bit more sensitive. It is concluded that some hybrid LSA models can generate more reliable LSIs compared to pure models. It is recommended that future studies employ, integrate, and compare both knowledge- and data-driven methods instead of using a single method for LSA.