Spatio-Temporal Analysis of Rainfall and Regional Groundwater Modelling of the West Coast Basins of India

Abstract

The Indian summer monsoon (June to September) is the backbone of the country's agriculture and allied sectors, exhibiting spatial and temporal variability across the Indian subcontinent. A growing body of research on climate change indicated the varying and patterns of the Indian summer monsoon. The Indian west coast is among the densely populated region with undulating topography. The region has the Western Ghats to the west and the Arabian sea to the east. Demands from agriculture and allied sectors in conjunction with the growing population have adversely affected the groundwater reserves in the region. The scenario has worsened owing to changing rainfall regimes over the recent decades. In this context, the present study aims at understanding the spatio-temporal patterns of rainfall and groundwater levels in the west coast basins of India. The feasibility of applying machine learning models in simulating the region's groundwater levels was also investigated. The west coast basins are categorised by the Central Water Commission (Ministry of Water Resources and Ganga Rejuvenation, Govt. of India) as sub-basin of Bhatsol and others, sub-basin of Vasishti and others, sub-basin of Netravati and others, sub- basin of Varrar and others and the sub-basin of Periyar and others. The departure analysis assessed the decadal variability and epochal behaviour of annual and seasonal rainfall in the west coast basins. The annual and seasonal rainfall departures displayed decadal variability among the west coast basins. The decades from 1980 to 1989 and from 2000 to 2009 were observed as the driest decade common among the west coast basins. An increase in dry rainfall-year frequency was noted after 1980 for Periyar, Varrar, and Netravati basins. Wavelet power spectrum analysis was conducted to examine the role of teleconnections in the region. Inter-annual periodicities of 2-4-years and 4-8-years were predominantly exhibited by Bhatsol, Vasishti, and Periyar basins with varying wavelet power for southwest monsoon rainfall. At the same time, Netravati and Varrar basins presented few short periodicities of 2-4-year band confined to early decades. However, statistically significant inter-decadal oscillations of 12-16-year period were evident among all the basins with moderate wavelet power. The inter-annual and inter-decadal variability in the distribution of southwest monsoon rainfall (or Indian summer monsoon) is clear from the periodicities obtained from the wavelet spectra. The widely used Sen's slope estimator investigated rainfall and groundwater level trends at annual and seasonal scales. The statistical significance was examined by the modified Mann-Kendall test (mMK). iSouthwest monsoon rainfall indicated a significant decline of -2.48mm/year, - 6.57mm/year, and -5.34mm/year at 5% significance level in the Periyar, Varrar, and Netravati basins respectively. The Vasishti basin indicated a significant increase of +5.89% in average winter rainfall per decade. The influence of parent distribution on the test power of the mMK test was analysed by Monte Carlo simulation of the generalised extreme value (GEV) distribution. The rank-based Mann-Kendall test may not attain the threshold power of 0.8 owing to heavy-tailed distribution or scale parameters. The trajectory of trends was extracted using the singular spectrum analysis (SSA) for rainfall and groundwater level time series. Significant trends by mMK for the monsoon and post-monsoon GWLs indicated a greater number of falls than rises. An average significant fall of -0.032m/year, i.e., a 7m decline in 12.3% wells was observed for monsoon, while -0.042m/year i.e., 0.92m fall in 11% wells, was indicated during the post-monsoon season. The artificial neural network (ANN) and support vector machines (SVM) were incorporated to examine the feasibility of machine learning models in predicting groundwater levels in the west coast basins. Type I models included abstraction and meteorological variables as predictors, while type II used only meteorological variables as predictors. The machine learning models, namely ANN1, ANN2, SVM1, and SVM2, performed well in predicting groundwater levels during the test period from 2012 to 2017. All four models showcased promising results for RMSE metrics with more than 93% of models in each category associated with an RMSE<0.2m. About 60.53% ANN1, 58.13% ANN2 models, 56.93% SVM1 models, and 55.26% SVM2 models exhibited an R2 >0.7, indicating the accountability of both ANN and SVM approach in predicting the groundwater levels (GWLs). The present study emphasizes the rainfall patterns on the west coast of India and the feasibility of using simple machine-learning algorithms to simulate groundwater level without an extensive hydrogeological dataset.