Time-series inversion of heavy metal and non-heavy metal water quality parameters and reveal their driving mechanisms using CNN-BiLSTM-Attention and PLS-SEM

Monitoring heavy metals and other pollution in drinking water sources is protecting public health, ensuring social stability, and promoting sustainable ecological development. Traditional fixed-point sampling approaches suffers from insufficient spatial coverage, weak remote sensing signals for heavy metals, and considerable challenges in achieving robust long-term inversion. To address these limitations, this study develops a multi-dimensional time-series feature fusion model (CNN-BiLSTM-Attention) for water quality parameter inversion. This model integrates a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism to enable monthly inversion of eight indicators, including dissolved oxygen (DO), chemical oxygen demand (COD), ammonia nitrogen (NH₃-N), fluoride (F − ), permanganate (COD M n ), zinc (Zn), selenium (Se), and mercury (Hg). This paper investigates the spatiotemporal evolution of both heavy metal and non-heavy metal water quality parameters during normal, dry and flood seasons. A partial least squares structural equation model (PLS-SEM) is constructed to quantify the direct and indirect effects of climate, human activities, and hydrological characteristics on the dynamic changes of the eight water quality parameters, thereby identifying their dominant drivers. The results showed that the CNN-BiLSTM-Attention model substantially improves inversion performance by capturing nonlinear spatiotemporal relationships among water quality parameters (mean R 2 = 0.791), representing a 32.05% increase over the conventional CNN-LSTM model (mean R 2 = 0.599). From 2018 to 2024, the overall water quality exhibited an improving trend, with notable decreases in Hg, Zn, Se, and COD concentrations (maximum reduction of 68.3%), while DO and NH 3 -N increased slightly. Spatially, pollutants accumulated at the reservoir edge and tributary confluences during the dry season and migrated toward the reservoir center in the flood season. Notably, the low-DO area largely overlapped with the peak COD Mn and COD region in the southwestern tributary of the reservoir. PLS-SEM revealed that the organic pollution load (COD Mn , COD) serves as the central hub driving heavy metal evolution, exhibiting a strong positive correlation with heavy metal indicators (Hg, Se; β = 0.946, P < 0.001). Human activities exert a direct negative effect on oxygen-consuming pollutants (DO, NH₃-N; β = −0.359, P < 0.01) and indirectly modulate heavy metal dynamics through the organic pollution load pathway ( β = −0.432, P < 0.05). Climatic factors (primarily rainfall, β = 0.534, P < 0.1) and water temperature ( β = −0.562, P < 0.05) constitute the dynamic basis of heavy metal evolution. The proposed time-series inversion and driving mechanism analysis method enables both temporal monitoring and attribution analysis of heavy metals and non-heavy metals in water bodies, providing reliable scientific support for water quality management in drinking water sources.