Monitoring and Temporal Prediction of Groundwater Quality using data Obtained via Probe sampling

Abstract

Despite advances in microelectronics and IoT (Internet of Things) that enable low-cost telemetry sensors, most environmental agencies still operate with periodic in situ sampling campaigns, characterized by sparse, irregular data and a high rate of missing values. While analytically reliable, this data was designed for post-facto regulatory diagnoses, not for continuous predictive systems or early warnings. This article presents a temporal predictive system for water quality based on non-telemetry data from multiparameter probes, applied to a 40-year historical series (1983–2023) of the industrial region of Cubatão (SP - Brazil). The LSTM, Random Forest, and XGBoost models are compared in predicting a new water quality index (IQA_New), sensitive to metals and industrial contaminants. The results show that LSTM consistently outperforms both Random Forest and XGBoost across all forecast horizons, achieving superior early warning capability (t+1: AUC 0.92; recall 0.74) while maintaining high sensitivity for the "Poor" class even at longer horizons (t+3: recall 0.69; AUC 0.89). The research contributes to fault imputation strategies in sensor series, temporal validation without data leakage (expanding window), and a modular index applicable to environmental instrumentation systems.