Integrating causal knowledge with satellite and glider data: a novel framework for marine water quality simulation

This study presents a causally informed machine-learning framework for jointly estimating coastal marine variables by incorporating prior structural assumptions into model training. The causal-informed neural network (CINN) uses an expert-specified directed acyclic graph (DAG) and model-based feature-perturbation effects as regularization targets; it is designed for prediction and does not identify causal effects from the observational data. The framework combines satellite remote sensing data with high-resolution in-situ glider measurements from the upper 0–2 m to jointly estimate temperature, salinity, dissolved oxygen, and chlorophyll-a. Under acquisition-date-blocked five-fold cross-validation, in which all observations from the same acquisition date were kept in the same fold, CINN achieved R² values of 0.961, 0.905, 0.898, and 0.725, respectively, and a uniform-average multi-output R² of 0.872. Regularization-weight, matched-input, and output ablations assess predictive sensitivity, while bootstrap intervals and Diebold–Mariano tests quantify uncertainty in the reported observation-level comparisons.