Development of a river dissolved oxygen prediction model integrating spatial effects and multiple deep learning algorithm

Dissolved oxygen (DO) in rivers serves as a vital indicator for assessing aquatic environmental health, as it significantly influences the survival of aquatic organisms and the stability of ecosystems. To address the nonlinear, complex, and periodic nature of DO time series, a novel prediction framework is proposed, in which Wavelet Convolution (WTConv), a technique traditionally used in image processing, is applied for the first time to DO forecasting. A hybrid architecture, named RIME-VMD-FECAM-WTConv-Transformer, integrates RIME (Rooted in Rime-Ice), Variational Mode Decomposition (VMD), Frequency Enhanced Channel Attention Mechanism (FECAM), and Transformer. In this model, the DO series is first decomposed into multiple sub-series via VMD, and each sub-series is individually trained and predicted using the FECAM-WTConv-Transformer structure; the final output is obtained by aggregating the predictions. The model is benchmarked against several baselines and evaluated using RMSE, MAE, R 2 , and PICP, which are recorded as 0.2935, 0.2053, 0.9852, and 98.9 %, respectively—outperforming all competing models. Cross-validation, significance testing, out-of-sample prediction, and Shapley Additive Explanations (SHAP) are also conducted to ensure the model's stability, reliability, interpretability, and performance. In addition, wavelet transform is used to explore the temporal correlations between DO and meteorological and water quality variables, further enhancing the interpretability of the deep learning approach. Overall, the proposed model demonstrates excellent predictive performance and holds strong potential for supporting the scientific management of surface water resources. • This study introduces a hybrid prediction model, named RIME-VMD-FECAM-WTConv-Transformer, for forecasting dissolved oxygen (DO) levels in rivers. This model incorporates multiple advanced techniques, including Variational Mode Decomposition (VMD), Wavelet Convolution (WTConv), Frequency Enhanced Channel Attention Mechanism (FECAM), and Transformer architecture, effectively addressing the nonlinearity, noise, and periodicity found in DO time series data. • WTConv, typically used in image processing, is applied for the first time to time-series prediction. This novel integration helps capture frequency-domain patterns, distinguishing between long-term trends and short-term fluctuations, while reducing model complexity by avoiding large convolution kernels. • The model utilizes VMD for decomposing DO series into multiple intrinsic mode functions (IMFs), allowing for improved signal representation and noise suppression. The multi-scale decomposition further enriches the model's ability to capture both local and global temporal features across different time frequencies. • The proposed model demonstrates outstanding predictive accuracy, outperforming several baseline models (including XGBoost and CNN-BiLSTM) in terms of RMSE (0.2935), MAE (0.2053), R 2 (0.9852), and PICP (98.9 %), thereby establishing its potential for environmental monitoring and management. • Incorporating data from surrounding monitoring stations (S1–S4) enhanced prediction performance. The study highlights the importance of spatial correlations in improving DO forecasting, with neighboring stations providing valuable anticipatory information for target locations.