Explainable soft-sensing of dissolved oxygen in a multi-station aquatic monitoring network using ensemble machine learning and neural networks

Dissolved oxygen (DO) is a critical water-quality indicator because it reflects aquatic system condition and directly informs monitoring and management decisions. This study develops and compares interpretable machine-learning and deep-learning models for multi-station DO soft-sensing from co-measured hydro-physical and water-quality variables. Model development was evaluated under a primary random split, a strict chronological train-past → test-future split, and leave-one-station-out (LOSO) validation to assess both temporal and spatial generalization. Baselines based on linear regression and seasonal mean prediction were also included. Among the evaluated models, ExtraTrees provided the strongest performance under the primary split (R² = 0.999706, RMSE = 0.020752 mg L⁻¹, MAE = 0.008211 mg L⁻¹), while chronological validation produced substantially lower but more realistic performance (R² = 0.539698, RMSE = 0.363559 mg L⁻¹, MAE = 0.268316 mg L⁻¹). Under LOSO validation, ExtraTrees retained comparatively strong station-transfer performance, with mean ± standard deviation values of R² = 0.9295 ± 0.1270, RMSE = 0.2432 ± 0.1436 mg L⁻¹, and MAE = 0.1591 ± 0.0950 mg L⁻¹. Ablation analysis showed that oxygen saturation was highly informative, but strong predictive performance remained possible even after its exclusion. SHAP interpretation consistently identified temperature as the dominant negative predictor of DO, with seasonal and salinity/conductivity effects providing secondary explanatory value. The present work contributes an interpretable and deployment-oriented framework for multi-station DO soft-sensing, in which validation design, target-linked predictor assessment, and feature explanation are integrated to support more defensible model selection and environmental monitoring applications