Input data uncertainty analysis with data augmentation methods for groundwater level prediction with the random forest model

Groundwater plays a crucial role in sustaining ecosystems and supporting agriculture and industry, particularly in arid and semi-arid regions. Accurate groundwater level (GWL) prediction is essential for sustainable water resource management. Artificial intelligence (AI)-based models have become powerful predictive tools, yet their reliability is often affected by input data uncertainty, arising from measurement errors, interpolation inaccuracies, and the inherent variability of hydrological processes. Addressing these uncertainties is critical to improving model performance and ensuring trustworthy predictions. This study aims to enhance the reliability of AI-based groundwater level prediction by analyzing and quantifying input data uncertainty using data augmentation techniques. Two data augmentation methods, including Jittering, Frequency Domain Transformation, were employed to assess the impact of input data uncertainty on model reliability within a Random Forest model for GWL prediction. These methods simulated inherent noise in input data, helping the model better adapt to real-world variations. The results demonstrated that the jittering method achieved a higher accuracy (R۲ = ۰.۸۶۱, NSE = ۰.۸۵۹), whereas the frequency-domain method produced narrower input data uncertainty bands (AW = ۰.۰۶۵, R-factor = ۰.۶۳۶). By evaluating and comparing the effectiveness of these techniques, this research provides insights into improving model robustness and reliable GWL prediction. The findings contribute to the development of more reliable AI-driven solutions for groundwater management, ultimately supporting informed decision-making and sustainable resource planning.

Keywords:

Groundwater level , Uncertainty analysis , Artificial intelligence model , Data Augmentation