A novel framework for river organic carbon retrieval through satellite data and machine learning

Rivers transport large amounts of carbon, serving as a critical link between terrestrial, coastal, and atmospheric biogeochemical cycles. However, our observations and understanding of long-term river carbon dynamics in large-scale remain limited. Integrating machine learning with remote sensing offers an effective approach for quantifying organic carbon (OC) from space. Here, we develop the Aquatic-Organic Carbon (Aqua-OC), a dynamic machine learning retrieval framework designed to estimate reach-scale river OC using nearly half a century of analysis-ready Landsat archives. We first integrate a globally representative river OC dataset, comprising 299,330 measurements of dissolved organic carbon (DOC) and 101,878 measurements of particulate organic carbon (POC). This dataset is then used to evaluate the performance of four machine learning methods, i. e., random forest (RF), extreme gradient boosting (XGBoost), Support vector regression (SVR), and deep neural network (DNN), using an optical water type classification strategy. We further leverage multimodal input features to enhance the Aqua-OC framework and OC retrieval accuracy by considering various factors related to OC sources and environmental conditions. The results demonstrate that the Aqua-OC can effectively estimate DOC (R2 = 0.68, RMSE = 2.88 mg/L, Bias = 2.63 %, Error = 12.52 %) and POC (R2 = 0.76, RMSE = 1.76 mg/L, Bias = 6.31 %, Error = 21.36 %). Additionally, the Mississippi River Basin case study demonstrates Aqua-OC's capability to map nearly four decades of reach-scale OC changes at a basin scale. This study provides a generalized method for satellite-based river OC retrieval at fine spatial and long-term temporal scales, thus offering an effective tool to quantify the rivers' role in the global carbon cycle.