Comparing Multisource Optical Satellite Imagery Products for River Discharge Estimation using Deep Learning

Remote sensing of river discharge (RSQ) has advanced rapidly with increasing satellite availability and the development of data-driven methods capable of generating predictions directly from images. Here, we retrain and evaluate the transferability of a recently developed transformer-based deep learning (DL) model with five medium resolution multispectral products using 2,760 gauges distributed globally. The model captures discharge-relevant signal directly from variations in reflectance. We further evaluated the impacts of sensor harmonization, atmospheric correction, and differences in revisit frequency. Performance was broadly consistent across products with median correlation and Kling-Gupta Efficiency (KGE) values ranging from 0.55 to 0.59 and -0.03 to 0.04, respectively. Differences between best and worst performing products per site generally varied moderately without exhibiting clear geographic patterns. Performance was largely insensitive to dominant land cover classes and dam presence but declined in correlation for rivers wider than 250 m. Gauge-calibrated, model-calibrated, and uncalibrated iterations of C/M approaches provide different baseline comparisons with respect to predictive skill and usage cases. Site-level scaling errors resulting in bias remains a primary limitation of the DL model. These findings demonstrate discharge-relevant information for image-based RSQ is transferable across optical sensors and processing levels, supporting flexible discharge monitoring globally using multisource satellite imagery.