Development of Predictive System for Avalanche Forecasting over North West Himalayas

Abstract

The research summarized in the work focuses on developing avalanche forecasting techniques tailored for the North-West Indian Himalayas by employing machine learning approaches and addressing critical data-related challenges, such as standardization, data quality, and missing values. To counter poor data quality at target locations, this study implements multivariate imputation by chained equations (MICE), successfully replicating snow events and weather variables, thus proving effective in data imputation. Additionally, the study introduces Boruta, a wrapper method based on Random Forest, for feature selection, which classifies variables into confirmed, tentative, and rejected categories based on their relevance to avalanche prediction, highlighting key parameters like storm snow, standing snow, and snow accumulation over 48 hours. HIM-Strat significantly improves snowpack analysis while reducing costs and risks associated with traditional snow stratigraphy, allowing for daily snowpack monitoring and enabling forecasters to enhance prediction accuracy. Region-specific models were developed for Stage-II in the Pir Panjal range and Drass in the Greater Himalayas, where HIM-Strat achieved a Heidke Skill Score (HSS) of 0.26 and 0.29, with an accuracy of 0.7 and 0.75, respectively. Beyond HIMStrat, additional numerical models such as the Hidden Markov Model for probabilistic analysis, the Nearest Neighbor Model for identifying past similar avalanche events, and an Artificial Neural Network to address data skewness were developed, culminating in an ensemble model combining all four approaches. This ensemble technique enhances predictive accuracy by leveraging the strengths of individual models, mitigating overfitting, improving robustness, reducing model-specific biases, and offering better generalization to unseen data.