Prediction of Groundwater from Principal Hydrological Elements Utilizing Vision Transformer

Abstract

Groundwater levels (GL) monitoring is an essential aspect of the hydrological cycle, and it serves as an indicator for various resource management methods in a watershed. Predominant techniques have shown that GL can be predicted for any geographic region if sufficient data exists for computation through sophisticated algorithms. One of the major issues in water governance is the lack of reliable and complete data for estimating and studying the GL decline trend. Hence, using reliable artificial intelligence (AI) models known for low informational needs and high prediction accuracy is more practical. This research aims to estimate groundwater prediction using a Vision Transformer (VTf) based on principal hydrological elements (PHE). PHE from 2017 to 2024 included flow level, velocity, rainfall, humidity level, average temperature, transpiration, and warm days. Traditional deep learning (DL) models need a fixed size of data, which greatly restricts their flexibility to analyze data of varying scales; hence, they fail to capture persistent dependencies. Overcoming these constraints is where VTf excels through the use of the Individual Attention Mechanism (IAM) of the transformer concept. VTf gives the best results when estimating GL to achieve the best accuracy.