An Innovative Ensemble Approach of Deep Learning Models with Soft Computing Techniques for GIS-based drought-zonation mapping in Rarh Region, West Bengal

Drought is a complex natural disaster impacting ecosystems and communities, making its identification crucial for mitigation efforts. This study aimed to assess drought scenarios in the Rarh Region of West Bengal at 3-, 6-, and 12-month intervals. The region is an amalgamation of the plateau and Gangetic delta, facing a decreasing rainfall trend, particularly in Birbhum and Purba Bardhhaman districts. Purba Bardhhaman, known for its good track of rice production, is now facing severe drought, which is a concerning matter. The study assessed their collinearity by evaluating 27 drought assessment variables grouped into meteorological, agricultural, hydrological, and socio-economic facets. A Multi-Layer Perceptron Neural Network (MLP NN) was applied as a benchmark, followed by a DenseNet neural network. Finally, a Hybrid Deep Learning Ensemble model was developed to compare precision and create a drought-prone map. Results indicated that, on average, 26.66% of the region is highly drought-prone at a 3-month interval, 20% at 6 months, and 25% at 12 months. The models were validated using ROC-AUC, Standard Error, and Asymptotic Significance. The Hybrid Deep Learning Ensemble model showed the highest accuracy, achieving 94.2%, 94.3%, and 95.3% at 3-, 6-, and 12-month intervals, respectively. This research provides valuable insights for policymakers in West Bengal to address the increasing drought risks in the region.