From Local Support to Spatial Extrapolation Delineating the Transferability Boundaries and Reliability of Remote-Sensing-Based TP Prediction in Urban River Networks

This dataset supports the evaluation of spatial transferability and prediction reliability for remote-sensing-derived total phosphorus (TP) in an urban river network in Baoshan District, Shanghai, China. It contains processed model inputs for 37,643 fine-scale water units, including 537 candidate predictor variables representing land use, potential pollution-source and human-activity proxies, bathymetry, and river geometry. The reference TP values are derived from ESA-MDN retrievals rather than independent in situ measurements. The dataset includes sample identifiers, TP responses, spatial cross-validation partitions, and heterogeneous graph inputs comprising six node types and thirteen directed relation types. Five-fold partitions are provided for spatial block lengths of 20, 50, 100, 200, and 400 m, with corresponding boundary buffers of 2, 5, 10, 20, and 40 m. Training, validation, testing, and exclusion roles are defined separately for each fold and spatial configuration; the exported training and testing folders therefore do not represent a single permanently disjoint split. These processed data support comparisons of local-support prediction and spatial extrapolation, assessment of performance decay with increasing spatial support distance, and analyses of predictive uncertainty and attribution reliability. Source-related predictors describe spatial associations and environmental-pressure proxies and should not be interpreted directly as pollutant discharge loads or causal effects.