A hybrid machine-learning model to map glacier-related debris flow susceptibility along Gyirong Zangbo watershed under the changing climate

Gyirong serves as an important channel to Chine-Nepal Economic Corridor, which is also the only land route for China-Nepal trade since the 2015 earthquake. However, the Gyirong corridor suffers from glacier-related debris flow from every April to September because of the complex topographic features and the changing climate. Therefore, a susceptibility map in response to precipitation and temperature change is timely, not only to ensure the safe operation of this corridor, but also to provide decision-makers a guidance for hazard mitigation and environmental remediation. Conventional method is difficult to consider and link the meteorological factors (e.g. temperature and precipitation), topographies, ecological, geological conditions all together to produce the susceptibility map, as such, machine learning is utilised to conduct the analysis. Logistic Regression (LR) and Support Vector Machine (SVM) were firstly applied to evaluate their efficiency and effectiveness of the performance of producing the susceptibility map. In order to improve the fitting and prediction accuracy (ACC), genetic algorithm - support vector machine (GA-SVM) and certainty factor - genetic algorithm - support vector machine (CF-GA-SVM) were conducted based on the initial analysis results of receiver operating characteristics curve (ROC) and ACC. Through the analysis, it can be seen that over 61% of the study areas have a high susceptibility to debris flow, requiring an intensive attention from the local government. To further optimise the computational time, when dealing with small amounts of sample data, SVM is more efficient than LR, but CF-GA-SVM can achieve the highest AUC (Area Under Curve) and ACC values, 0.945 and 0.800, respectively. Overall, CF-GA-SVM model presents a relatively high robustness according to sensitivity analysis