Groundwater is the most precious natural resource in the Gaza Strip, South Palestine. The only source of groundwater supply in the area is the Mediterranean coastal aquifer of the Gaza Strip. However, due to a large population growth in recent decades, with an ever-increasing demand for domestic and agricultural water, groundwater in the region has been overexploited over the years. This has led to excessive reductions in yields, deterioration of ground water quality and pumping wells going dry. Therefore, for maintaining the sustainability of the Gaza groundwater system and to forestall imminent future problems, a better understanding of its dynamics is needed. To do this properly, numerical groundwater modeling must be done.
In the present study artificial neural networks (ANN) is applied as a new approach for groundwater management in the Gaza coastal aquifer, for the purpose to investigate the effects of hydrological, meteorological and human factors on the dynamic groundwater levels in the aquifer. The initial ANN model for predicting groundwater levels is set up using monthly groundwater time series data recorded between 2000 and 2010 at 70 wells across the Gaza Strip and employing seven independent predictor variables, namely, initial groundwater level, abstraction rate, recharge from rainfall, hydraulic conductivity, distance of the pumping wells from the coastal shoreline, depth to the well screen and well density. The best architecture of this initial ANN model found by trial and error turns out to be a 3-layer perceptron network (MLP), i.e. is an ANN with one hidden layer between input and output layer.
However the subsequent sensitivity analysis of this initial ANN model shows – from the computation of the ratio of the mean square error without a particular variable included to that of the full model – that two of the seven input variables are non-influential for the water level predictions and can thus be discarded from the ANN model. The latter is then revised accordingly and the final ANN model obtained again after numerous trials is a 4-layer MLP, with an input layer consisting of five neurons, a first hidden layer with 30 neurons, a second hidden layer with 20 neurons, and the output layer with one neuron. Finally, in order to get some more physical insight into the aquifer system’s behavior, response graphs and response surfaces are visualized which indicate, among others, that the final water levels are positively correlated with the initial water levels and with the groundwater recharge and negatively with the pumping rate, whereas their dependencies on the well screen depth and on the well density are somewhat ambiguous.