Constructing the machine learning techniques based spatial drought vulnerability index in Karnataka state of India

• Published in: TERI information digest on energy and environment Vol. 20; no. 3; pp. 394 - 395
• Main Authors: Saha, S, Gogoi, P, Gayen, A, Paul, G C
• Format: Journal Article
• Language: English
• Published: New Delhi The Energy and Resources Institute 01.09.2021
• Subjects: Agricultural ecosystems; Agricultural production; Algorithms; Anthropogenic factors; Artificial neural networks; Bagging; Drought; Groundwater; Groundwater storage; Human influences; Learning algorithms; Machine learning; Mathematical models; Modernization; Natural resource management; Natural resources; Neural networks; Overexploitation; Parameter sensitivity; Rank tests; Resilience; Root-mean-square errors; Water availability
• ISSN: 0972-6721; 1875-9297

The drought-induced vulnerability is owed much to rapid modernization, climate extremes, and overexploitation of natural resources. However, the drought, a natural phenomenon, is amplified by anthropogenic activities and in its enormity influences water availability, agricultural productivity, ecosystem, and groundwater storage. Karnataka is frequently affected by the drought that causes a huge loss in the agricultural sector and other allied sectors. Therefore, it is essential to measure the vulnerability status for the better management of natural resources in the state of Karnataka. No advanced models are being used yet to portray the drought vulnerability status. Different advanced machine learning models are effective in predicting various physical vulnerabilities. The aim of this study was to use sophisticated machine learning models to precisely define relative drought vulnerability. In that endeavour, it used two advanced machine learning algorithms (MLAs), namely, bagging and artificial neural network (ANN), which are still not used in this field. Twentysix meteorological and socio-economical parameters were considered to find the most drought vulnerable areas. The predisposing parameters were classified as resilience (7 parameters), sensitivity (9 parameters), and exposure (10 parameters). The researchers have produced drought vulnerability maps for overall condition, resilience, sensitivity, and exposure. The relative drought vulnerability maps (RDVMs) clearly show that 40.87-52.03% of areas fall under very high vulnerability, situated in the central and eastern parts of the state. The prediction capacity of newly built models was judged for efficiency, root mean square error (RMSE), true skill statistics (TSS), Friedman and Wilcoxon rank test, and the area under the curve (AUC) of the receiver operating characteristic (ROC). All of them showed satisfactory results - the RMSE values of 0.32 and 0.33, TSS values of 0.82 and 0.81, and AUC values of 86.50% and 84.20% as obtained by ANN and bagging models, respectively. The produced RDVMs demonstrate the urgency of policy interventions to minimize vulnerability in prioritized areas. (10 Figures, 10 Tables, 97 References)