Short-term forecasting of daily dissolved oxygen in streams using SWAT, remote sensing and explainable machine learning

Dissolved oxygen (DO) concentration is a key indicator of aquatic ecosystem health yet modelling and forecasting it accurately at daily resolution remains challenging due to the complex interplay of influencing factors. In this study, we explore the potential of combining diverse data sources, including in-situ observations, satellite-derived time series, and modelled catchment flow and nutrient loads from SWAT models, to forecast daily DO concentrations in streams. This is the first study that integrates SWAT-simulated hydrological and nutrient outputs with remote sensing data and machine learning to forecast daily DO in streams, advancing beyond previous models that rely solely on observations or use models only for DO simulation. We evaluated the performance of three widely used machine learning and deep learning models (Random Forest, LSTM, and Transformer) at three stream sites in the Upper Piako catchment in New Zealand. The results showed that model performance varied by site, emphasizing the importance of input variables tailored to local conditions. While water temperature consistently emerged as a dominant predictor, other variables such as baseflow and vegetation indices also were found to be important predictors. Our findings highlight the importance of integrating domain knowledge to guide feature selection, particularly when combining observational, remote sensing, and modelled data, to improve DO forecasting accuracy.