Code and derived remote-sensing inputs for two-stage monthly river-discharge reconstruction and source-station transfer experiments

This archive provides research code and derived remote-sensing inputs for a two-stage monthly river-discharge reconstruction framework evaluated at 16 river reaches in Northeast China, together with experiments using additional source stations from the United States.

The archive includes: (1) portable main-model inference code and trained model parameters; (2) derived monthly river widths, satellite precipitation, station descriptors, and observation-support information; (3) four locally adapted comparison methods, comprising linear regression, power-law regression, a rank-based lookup table, and stochastic quantile mapping; (4) supplementary direct-MLP, no-dynamic-rain, and frozen-parameter smoothing ON/OFF experiments; and (5) complete-station leave-one-out workflows for Northeast China, source-extension experiments with 4–10 US stations and ten compositions per source count, and the additional 11-US-station experiment.

No observed discharge/runoff records or ground-observed water-level records are included. Pointwise observation errors and predicted-to-observed discharge ratios are also excluded.

Main-model inference can be reproduced using the supplied inputs and parameters. Retraining and independent recomputation of observation-based evaluation metrics require separately authorized ground observations and historical input registries that are not distributed in this archive. The four literature-based comparison methods are local adaptations rather than complete reproductions of the original authors’ processing chains. The 11-US-station experiment contains one source composition and does not provide a ten-composition uncertainty interval.