Ground Water Level Prediction with Non-Continuous Unevenly Spaced Small Timeseries using Machine Learning

Ground water (GW) level prediction is an invaluable tool in GW management. Even with dense network of 256 GW monitoring wells at National Capital Territory of India, the GW level prediction possesses a significant challenge not only because of its small data size (${1 0} {- 4 8}$ observations per well during 2005-2020) but also due to unevenness and non-continuity between two observations. With the motive to aid local level GW management, we model each well independently with Support Vector Regression (SVR) by a specifically designed model selection strategy that uses the Corrected Akaike Information Criterion to choose the model that generalizes effectively. Of the 80 wells with sufficient data for modeling, 60 observed wells returned good results when evaluated using mean absolute errors (MAE) ranging from 0.2m to 0.8m below ground level (mbgl). This study demonstrated this framework is an efficient tool for prediction using small datasets with significant gaps.