An optimization on machine learning algorithms for mapping snow avalanche susceptibility

Abstract

Mapping avalanche-prone areas to mitigate damages is important and vital for safety and development planning. New hybrid models are introduced for snow avalanche susceptibil- ity mapping (SASM) in the Zarrinehroud and Darvan watersheds in northwestern Iran. A hybrid of four learning models—radial basis function, multi-layer perceptron, fuzzy ART- MAP (or predictive adaptive resonance theory (ART), and self-organizing map (SOM)— with three statistical algorithms—frequency ratio, statistical index, and weights-of-evi- dence—and K-means clustering integrated 20 factors and 177 avalanche locations. The areas most likely to produce snow avalanches were identified. The relative importance of the predictive factors was determined by analyzing the information gain ratio (IGR). Slope (average merit (AM) = 0.48055) and LS(AM = 0.00202) were the most and least impor- tant factors. Positive predictive value, negative predictive value, sensitivity, specificity, area under the curve (AUC), standard error (SE), mean square error, and root mean square error (RMSE) were used to validate the results of the models. The K-means-SOM hybrid model (AUC = 0.811, SE = 0.0548, RMSE = 0.39005) produced the best results of the hybrid models. This study demonstrates that SASM can help local managers and planners mitigate losses of life and damages caused by avalanches.