Spatiotemporal Patterns and Multiscale Drivers of Riverine Water Quality in China: A Continental-Scale Analysis via Physics-Informed Deep Learning

The growing availability of large-scale environmental datasets offers a foundational opportunity to decode continental water quality dynamics; however, disentangling the scale-dependent interplay between transient meteorological forcings and persistent catchment attributes remains a challenge. Focusing on the pronounced hydro-climatic gradients of mainland China, this study presents a comprehensive synthesis of heterogeneous environmental data to bridge the gap between data-driven predictability and mechanistic understanding. We curated an extensive dataset anchored by 4,077 monitoring sections across nine major river basins, incorporating continuous daily observations from April 2014 to February 2025. This dataset covers 10 critical water quality parameters, encompassing a broad spectrum of physicochemical, nutrient, and biological indices.To systematically characterize driving mechanisms, we structured a multi-source explanatory framework that explicitly partitions predictors into two categories: 12 dynamic time-varying forcings (capturing transient fluctuations via meteorological variables and hydrological fluxes) and over 30 static attributes (representing physiographic contexts and anthropogenic footprints, such as land use intensity and reservoir regulation). To decode the nonlinear dynamics of this high-dimensional system, we propose a Physics-Data Coupled Framework employing an Encoder-Decoder architecture integrated with Multivariate Singular Spectrum Analysis. A key innovation is the embedding of explicit physical mechanism gates within the network, designed to ensure hydrological and biogeochemical consistency in deep learning predictions. Beyond enabling robust long-term forecasting, this framework facilitates multi-scale interpretability, allowing for the assessment of how dominant drivers shift across weekly, monthly, and annual resolutions. The analysis elucidates the differential roles of meteorological events in modulating high-frequency variability versus static landscape features in defining long-term baselines, offering a consistent methodological paradigm to support decision-making under changing environmental conditions.