Manifesting deep learning algorithms for developing drought vulnerability index in monsoon climate dominant region of West Bengal, India

• Published in: Theoretical and applied climatology 2023, Vol.151 (1-2), p.891-913
• Main Authors: Saha, Sunil, Kundu, Barnali, Saha, Anik, Mukherjee, Kaustuv, Pradhan, Biswajeet
• Format: Article
• Language: English
• Subjects: Accuracy; Agricultural drought; Algorithms; Analysis; Aquatic Pollution; Artificial neural networks; Atmospheric Protection/Air Quality Control/Air Pollution; Atmospheric Sciences; Climate; Climate science; Climatic indexes; Climatology; Data mining; Deep learning; Drought; Drought index; Earth and Environmental Science; Earth Sciences; Economic models; Economics; Environmental hazards; Error analysis; Gaussian process; Hydrologic drought; Hydrology; Learning algorithms; Machine learning; Monsoon climates; Monsoons; Neural networks; Performance indices; Root-mean-square errors; Socioeconomic aspects; Socioeconomics; Vulnerability; Waste Water Technology; Water Management; Water Pollution Control; Weather; Weather forecasting
• ISSN: 0177-798X; 1434-4483
• DOI: 10.1007/s00704-022-04300-4

Drought is a natural and complex climatic hazard that has consequences for both natural and socio-economic contexts. The drought vulnerability assessment is an important contribution towards reducing its impact on the society and people. Although there has been significant research progress in the forecasting and prediction of droughts, but its vulnerability evaluation using high-precision deep learning algorithms has not yet been made. Keeping these things in mind, the present research is focused on preparing drought vulnerability map (DVM) for the monsoon climate dominant region of West Bengal, India. Fifty drought vulnerability determining variables were employed and divided into four categories based on the drought types: hydrological drought, agricultural drought, meteorological drought, and socio-economic drought. Then one deep learning model, i.e., Deep Learning Neural Network (DLNN) and two machine learning methods, i.e., Artificial Neural Network (ANN) and Multitask Gaussian Process (MGP) were used for preparing the drought vulnerability maps (DVMs). Results show nearly 24% of the study area is very highly vulnerable to drought. Several statistical performance indices such as the receiver operating characteristic (ROC) curve, mean-absolute-error (MAE), root-mean-square-error (RMSE), accuracy, Jaccard, F1 measure, MCC, and K-index were applied for judging the efficacy of the DVMs. The results show that the DLNN performed best, with an AUC of 94.8%, followed by ANN (89.66%) and MGP (87.08%). Hence, the deep learning algorithms performed better than the traditional machine learning models to prepare the DVMs. The combination of various parameters of different types of droughts can accurately predict the drought vulnerable zones with very high accuracy. This can be useful for policy-makers in managing drought severity in the south-western part of West Bengal.