Predicting heterogeneous and non-linear processes remains a fundamental challenge in Earth sciences. Here, we present an artificial intelligence (AI)-guided, closed-loop workflow that iteratively combines predictive modeling with targeted field sampling to rapidly improve environmental predictions. We demonstrate our workflow by predicting oxygen consumption, a key process in rivers, across the contiguous United States (CONUS). Our approach consists of multiple iterative loops of measurements and models, combining distributed participatory field sampling, lab analysis, automated machine learning (ML) predictions, and error and distinctiveness analyses to autonomously guide the next sampling at optimal site locations. Specifically, we used an ensemble of stacked regressors to both maximize predictive performance and quantify predictive uncertainty, which then informed the data acquisition policy. Over 18 iterative loops, we increased the predictive power of sediment oxygen consumption across CONUS rivers by over fifteenfold between the first and last iteration. Relative to our last sampling iteration, our first sampling missed rivers with outlier values and underestimated median oxygen consumption rates by 68%. In addition to identifying areas of high oxygen consumption rates, iterations enabled refinement of laboratory and data handling methods, and engagement with a broad community of field researchers. We conclude that AI-guided iterative loops between targeted sampling and predictive modeling are a powerful and efficient approach for improving predictions of heterogeneous environmental processes. The architecture of our workflow balances model flexibility with interpretability and can be readily applied to other domains where uncertainty- and distinctiveness-aware site selection is crucial.