Water RSNet: An improved water quality prediction model combining dual-temporal and dynamic multimodal fusion

Accurate lake water-quality forecasting is limited by insufficient modeling of interactions among environmental variables, inadequate extraction of periodic frequency-domain patterns, background interference in remote-sensing imagery, and inflexible multimodal fusion. To address these limitations, this study proposes Water RSNet, which integrates water-quality, hydrological, meteorological, atmospheric, and remote-sensing data. First, TC-Adapformer adds a variable-channel attention path parallel to the temporal path to jointly capture temporal dependencies and dynamic cross-variable interactions. Second, FC-FEDformer retains seasonal-trend decomposition and frequency-enhanced modeling while introducing a channel-interaction path to represent periodic patterns and frequency-scale variable coupling. Third, spatial residual convolution combines spatial attention with residual connections to suppress irrelevant backgrounds and preserve water-body boundaries and local high-response features. Fourth, a two-stage gated fusion mechanism with cross-attention dynamically weights remote-sensing submodalities and fuses remote-sensing and time-series representations according to sample conditions. On the Guanting Reservoir dataset, Water RSNet achieved overall RMSE, MAE, and R² values of 0.10909, 0.06639, and 0.99289, respectively, outperforming the comparison models. Ablation and multi-timescale evaluations verified the effectiveness of the four modules and the stability of time-frequency modeling. External validation using Miyun Reservoir data yielded RMSE, MAE, and R² values of 0.02273, 0.01816, and 0.98425, respectively, demonstrating applicability in another reservoir and across different years. These results show that Water RSNet provides an effective framework for multi-indicator lake water-quality forecasting.