<p>Snow avalanches pose a significant threat to both individuals and infrastructure. Deep learning algorithms have been shown to be an efficient tool for modeling snow avalanche and other similar natural disasters, but they require a large sample size for training. However, some regions do not have availability to the required amount of data. This study utilizes established techniques and approaches to address this shortcoming so that these advanced algorithms can be applied even in regions with limited data. It utilizes the recurrent neural network algorithm to model snow avalanche susceptibility, applies a robustness maximization approach to prevent overfitting, and uses three meta-heuristic algorithms for hyperparameter optimization: grey wolf optimizer, particle swarm optimizer, and artificial bee colony optimizer. A performance comparison with other models, including deep neural networks and support vector machines, using the same training strategy, revealed that optimized recurrent neural network models are significantly better suited for datasets with limited sample sizes. The RNN-ABC model demonstrated superior predictive performance (AUC = 0.9710, accuracy = 0.9318, RMSE = 0.2354, sensitivity = 0.9090, and specificity = 0.9545) compared to the RNN-PSO and RNN-GWO models. Relief-F variable importance analysis identified lithology, aspect, land use, slope position, and proximity to streams and roads as key factors in this region. The designed process shows significant effectiveness in regions with limited data size and quality. This hybrid approach can theoretically be applied to many different regions with data scarcity, and possibly even for other natural hazards, providing significant prediction reliability improvement over previous methodologies.</p>