Proposing an ensemble machine learning based drought vulnerability index using M5P, dagging, random sub-space and rotation forest models

Drought is one of the major barriers to the socio-economic development of a region. To manage and reduce the impact of drought, drought vulnerability modelling is important. The use of an ensemble machine learning technique i.e. M5P, M5P -Dagging, M5P-Random SubSpace (RSS) and M5P-rotation forest (RTF) to assess the drought vulnerability maps (DVMs) for the state of Odisha in India was proposed for the first time. A total of 248 drought-prone villages (samples) and 53 drought vulnerability indicators (DVIs) under exposure (28), sensitivity (15) and adaptive capacity (10) were used to produce the DVMs. Out of the total samples, 70% were used for training the models and 30% were used for validating the models. Finally, the DVMs were authenticated by the area under curve (AUC) of receiver operating characteristics, precision, mean-absolute-error, root-mean-square-error, K-index and Friedman and Wilcoxon rank test. Nearly 37.9% of the research region exhibited a very high to high vulnerability to drought. All the models had the capability to model the drought vulnerability. As per the Friedman and Wilcoxon rank test, significant differences occurred among the output of the ensemble models. The accuracy of the M5P base classifier improved after ensemble with RSS and RTF meta classifiers but reduced with Dagging. According to the validation statistics, M5P-RFT model achieved the highest accuracy in modelling the drought vulnerability with an AUC of 0.901. The prepared model would help planners and decision-makers to formulate strategies for reducing the damage of drought.