Artificial intelligence-based approach to study the impact of climate change and human interventions on groundwater fluctuations

Groundwater resources accomplish one-third of Earth's freshwater demand (Brar et al. 2016). India uses an estimated amount of 230 km³ of groundwater per year, making the country the leading consumer of groundwater in the world (Das & Burke 2013). Among all states of India, Punjab covers around 1.53% of India's geographical area. Punjab's climate is semi-arid (Dhillon et al. 2019). Agriculture in Punjab is highly intensive with heavy requirement for water which cannot be met by rainfall. In addition, surface water resources are fully utilized through canal irrigation systems and existing surface water resources are unable to meet the requirements of agriculture. Therefore, groundwater is used as the primary source of irrigation water to grow crops, mainly paddy and wheat.

The average water table depth was found to be declining at a rate of 47.6 cm/year during 1998–2013 (Brar et al. 2016). Analysis of 138 blocks revealed that 109 are overexploited, 2 critical, 5 semi-critical, and only 22 are safe blocks (Anon 2018).

The decisions regarding water resource management need accurate information for groundwater level. This useful information can be provided by groundwater level modeling to policy-makers to get suitable outcomes. Multiple linear regression (MLR) is a linear approach for modeling groundwater and in recent studies this method has been applied to model and analyze groundwater recharge. MLR typically estimates the level of correlation between two or more predictors (independent variable) and one response variable (dependent variable). Mogaji et al. (2015) evaluated and forecasted groundwater recharge rates using the MLR technique in the southern part of Perak, Malaysia. It was concluded that the developed MLR model can be used in any place with the same geology. The simulation of significant amounts of data in hydrological and groundwater systems contain non-linear relationships between variables (Osman et al. 2021). The limitations of non-linearity can be overcome by artificial neural network (ANN) models, which have been used by many researchers (Lai et al. 2019). ANNs can offer a more active strategy for forecasting groundwater amounts in a vigorous and highly unreliable system. ANN consists of three layers, namely input, hidden, and output. The total number of input variables represent the total number of neurons in the input layer. The hidden layer is connected to the input layer and the output layer is connected to the hidden layer.

Mohanty et al. (2015) simulated groundwater level fluctuations weekly in multiple wells located over a river basin in the Mahanadi Delta of Odisha, India. Groundwater level was forecasted 1 week in advance with a reasonable accuracy by the developed ANN model.

There is a need to develop a prediction model to study groundwater fluctuation so that the effect of sustainable use of water resources on groundwater can be studied. The main objective of this research is to develop a groundwater model in Central Punjab during 1998–2019. The aims of the present research are: (1) to evaluate and confer the effect of climatic factors and human interventions on groundwater depth and (2) to observe the main factors responsible for groundwater depth fluctuations.

Broadly, Punjab can be classified into three zones, North-East, South-West, and Central. The Central Punjab comprises 10 districts as mentioned in Table 1. Central Punjab represents an area of 18,000 km² which covers more than 36% of the total area of Punjab. Their geographic description is given in Table 1. Keeping in view the problems of groundwater depth decline in Central Punjab, a trend analysis of 10 districts of Central Punjab was carried out.

In the present study, important factors responsible for groundwater fluctuations in Central Punjab were used. Series of annual rainfall, paddy area, basmati area, sugarcane area, non-paddy area, canal-irrigated area, tubewell density, and groundwater depth were examined. The gridded data of rainfall from 1998 to 2019 for different districts of Punjab were collected from the India Meteorological Department, New Delhi and analyzed on an annual basis (Pai et al. 2014). The paddy area, basmati area, sugarcane area, non-paddy area, canal-irrigated area, and tubewell density were gathered from statistical abstracts (Statistical Abstract of Punjab 2017, 2019). Initially, the water table for various districts was computed using the Kriging technique described by Kaur et al. (2012). To study the behavior of groundwater, data on groundwater levels during the pre-monsoon month (June) was obtained from the Directorate of Water Resource & Environment and Directorate of Agriculture for the study period (Singla et al. 2022).

At the beginning of the training, network weights and biases were assigned randomly. The solver for weight optimization is a stochastic gradient-based optimizer (Adam). The algorithm then projected these forwards from the input layer to the hidden layer. The results obtained from the hidden layer were propagated to the output layer. Then the error was calculated between the value of the output layer and the observed groundwater recharge in training data. The connection weights and biases were automatically adjusted until the network error reached a predetermined value by iteratively propagating the error back to the network.

The datasets of rainfall, sugarcane area, rice area, basmati area, non-paddy area, canal-irrigated area, and tubewell density were used as input and the values of groundwater depth was the output target variable. The input variables and the target output variable were standardized before the training and testing.

The MLR model and ANN model were built using the Python packages StatsModels and scikit-learn, respectively. Tensorflow, a Python-based end-to-end machine learning platform, was used to create the model. All of the modeling, data analysis, and visualization were done in the Python 3.5 environment. Python was chosen for the work environment because of its cross-platform processing capability, open-source code, presence of third-party modules, and extensive support libraries.

The larger the network error after the input variable is excluded compared to the primary error (for a network with all the input variables), the more sensitive the network is to the lack of this variable.

Sensitivity analysis of ANN for groundwater depth was exhibited as discussed in Section 2.8 (Table 5). After eliminating each indicator one by one and running ANN with six indicators and new error is determined for each eliminated variable. Then ratios (W) of new error and original error are calculated.

The groundwater depth is most affected by tubewell density followed by rainfall, area under rice, and sugarcane respectively (Table 5). Canal-irrigated area has the lowest effect on the output. In general, it can be revealed that the importance of input parameters of the model can be differentiated with the use of sensitivity analysis for the ANN model. In addition to this, it should be observed that this is immensely helpful when there are various input variables for the ANN model. From the sensitivity analysis, it is noticed that tubewell density, rainfall and rice area had a major influence on groundwater depth.

Groundwater declination is one of the major concerns for Central Punjab, as groundwater has fallen around >1 m depth annually over the last 20 years (Singla et al. 2022). The ANN and MLR models were an approach to making a forecasting model for measuring groundwater depth, as groundwater depth is influenced by both climatic factors and human interventions. So, models were developed keeping in view the influence of both factors. ANN performed better than MLR in the present study.

The sensitivity analysis showed that groundwater is greatly influenced by pumping and rainfall. This is reasonable, since more pumping is projected to increase groundwater withdrawal. The rice area is also found to be a significant influencer. The area under is approximately 40 times more than the sugarcane area (Statistical Abstract of Punjab 2017, 2019). The deficiency in rainfall over the years consequently impact evapotranspiration rates and subsequently the water requirements. Thus, the crops will need to take up more water from the groundwater. Rice area was found to be another major influencer on groundwater depth. Water productivity of rice (quantity of water required to produce 1-kg rice) in the state in the triennium (TE) ending 2013–14 was 5,337 l, whereas the all-India average was 3,875 l (Central Ground Water Board 2018). This is also due to applying a higher number of irrigations than the recommended doses. The paddy area in Central Punjab increased by 0.23 million ha during 1998–2017, indicating water requirement of (3.43 BCM (billion cubic metres)), considering an average of 1,500 mm irrigation water (Singla et al. 2022).

The decline in water table depth can only be addressed by reducing demand on the groundwater table. It is only possible with crop diversification (Kaur & Vatta 2015). Central Punjab needs to shift a huge area from under the paddy. It is of utmost importance for ensuring sustainable agriculture, ensuring livelihoods, saving water for future generations and saving Punjab from the looming desertification.

The present study was done to develop a prediction model using different techniques to study groundwater fluctuations in Central Punjab. Climatic factors like rainfall and other factors like area under rice, basmati, non-paddy, sugarcane, tubewell density, and canal-irrigated area were considered as input. Both ANN and MLR techniques were used to develop the model and ANN performed better. Furthermore, sensitivity analysis revealed that rainfall, rice area, and tubewell density are the main factors that cause groundwater depletion. Furthermore, the improvements in the model can be made by taking into consideration micro scale/district level analysis. Central Punjab needs huge diversification in rice for addressing the groundwater issue. The main function of the model is to study groundwater fluctuation, when diversified rice is considered.

All relevant data are available from an online repository or repositories.

The authors declare there is no conflict.