Multimodal remote sensing and meteorological estimation of long-term spatial-scale non-point source pollution in inland reservoir

• Multimodal model of remote sensing imagery, meteorological data and lag effects. • 30-year, spatial assessment of NPS pollution across 7 water quality parameters. • Multimodal models achieved higher accuracy and broader parameter estimation. • Meteorological lag effects greatly enhanced model accuracy and interpretability. Non-point source (NPS) pollution is one of the major threats to inland waters. This issue is particularly pronounced in remote regions, where the scarcity of long-term and spatial-scale water quality data hampers effective NPS pollution assessment. This study selected Chitian Reservoir, a tropical drinking water source, as a case study. A total of 323 samples were obtained from 2021 to 2023. Water quality observations were used as labels, and Landsat imagery together with meteorological factors with up to a two-day lag were used as input features. Based on four machine learning algorithms, multimodal water quality models with both spatial and temporal attributes were developed. Results showed that the multimodal models effectively estimated the concentrations of multiple water quality parameters, including total nitrogen (TN), total phosphorus (TP), ammonium nitrogen (AN), chlorophyll-a (CA), permanganate index (PI), total suspended substance (TSS), and turbidity (TU), achieving R 2 values of 0.71, 0.50, 0.70, 0.53, 0.83, 0.80 and 0.65, respectively, on the test set. Compared to unimodal models based on either remote sensing or meteorological factors, the multimodal models could estimate 3–4 additional water quality parameters, with accuracy improvements ranging from 2.01% to 63.80%. SHAP analysis revealed that spectral bands and lagged meteorological factors contributed significantly to the estimation of water quality parameters. When the models were extended to the period from 1994 to 2023, the spatiotemporal characteristics of water quality parameters exhibited significant differences. AN and TSS concentrations were higher in nearshore areas, while other parameters were higher in central zones. TP showed a clear declining trend, whereas other parameters remained relatively stable over the three decades. Concentrations of TP, AN, PI, TSS, and TU showed significantly higher during the rainy season, while TN exhibited the opposite trend. The proposed multimodal water quality model substantially improved both the number and estimation accuracy of water quality parameters. It identified the sensitive areas and key periods of NPS pollution in Chitian Reservoir across multiple dimensions, including a 30-year temporal span, spatial distribution, and multiple water quality parameters. These findings provide scientific support for water environment management in the study watershed and offer a valuable technical reference for long-term, spatial-scale, and multi-parameter water quality monitoring and NPS pollution control worldwide.