Groundwater quality assessment using few-shot learning with prototypical, Siamese, and matching networks

Groundwater quality assessment is critical for sustainable water resource management, especially in hyperarid regions like Adrar, Algeria, where data scarcity and environmental challenges hinder traditional monitoring methods. This study explores the application of three Few-Shot Learning (FSL) algorithms: Prototypical Networks, Siamese Networks, and Matching Networks, for groundwater quality classification using limited datasets. The dataset comprises 166 groundwater samples from the Adrar region, characterized by five quality classes: “Excellent,” “Very Good,” “Good,” “Satisfactory,” and “Unsatisfactory.” Results demonstrate that Prototypical Networks outperform other FSL algorithms, achieving 93 % accuracy with 10 support samples per class, while Siamese and Matching Networks achieve 90 % and 88 % accuracy, respectively. The study highlights the potential of FSL in addressing data scarcity, offering a cost-effective and efficient approach for groundwater quality assessment in data-scarce regions. The findings underscore the importance of FSL in complementing traditional methods, particularly in hyper arid areas where data collection is challenging. • Comparative analysis of Prototypical, Siamese, and Matching Networks for groundwater quality. • Application of few-shot learning to assess groundwater quality with limited data. • Evaluates model performance in classifying groundwater contamination levels. • Highlights advantages and challenges of deep learning in hydrogeological studies. • Provides insights for improving AI-driven groundwater quality monitoring systems.