The purpose of this research is to find the most effective way to forecast groundwater levels for wells in the Mesilla aquifer region of New Mexico. Although water level forecasting has many uses and applications, it is particularly useful in New Mexico. In 2015, New Mexico was home to 24,000 farms spanning 43.2 million acres. With more than 50% of the state's land being dedicated to agriculture, predicting groundwater levels allows municipalities to plan in terms of water conservation and allocation throughout residential and agricultural areas. In this work, we particularly focus on the Mesilla aquifer region of southern New Mexico, which is a transboundary aquifer between the US and Mexico under an intensive water use for agriculture and human consumption. 
Recently, machine learning-related approaches have also shown great predictive capabilities in the field of water resources. Several works in this field have used state-of-the-art techniques based on Recurrent Neural Networks (RNNs) to predict groundwater levels by using historical data, such as groundwater levels and other weather-related attributes (e.g. Precipitation, temperature, etc.). We observe that the wells in the Mesilla aquifer region are uniquely positioned, and hence propose a machine learning technique that is enhanced using spatial analysis.
In this research work, we focused on using the state-of-the-art Recurrent Neural Networks called Gated Recurrent Units (GRU). We noticed better accuracy by using GRU instead of Long Short Term Memory (LSTM). We created a baseline model (called Model1) using a groundwater well with water levels dating back 775 months. This model was built with GRU using the Adam optimizer. Further, we included additional weather-related data, such as precipitation, temperature, depth from the top of the well to the water surface, and land surface elevation in the training model (called Model2). Finally, we performed spatial analysis on each well to improve the predictions of the previous models. With over 300 wells in the data, we  sorted the distance of each well from the target well. We used the data from neighboring wells while training our model. After experimenting on different parameters (number of neighboring wells, amount of historical data to consider, number of weather-related attributes to choose from), we achieved the best results after using the data from five nearest wells, limiting the historical data to 10 years, and by using historical precipitation and temperature data along with the groundwater levels. We improved the accuracy (measured in terms of RMSE) of the baseline model by about 80% by integrating weather-related information and data from neighboring wells into the training model. For future work, we will focus on comparing our results with those using classical methods such as the modular finite-difference flow model (Modflow).