We propose an integrated framework for water contamination detection that combines physics-informed deep learning and adaptive management. The core of the system is AquaDynNet, a multi-branch convolutional neural network designed to simulate the spatiotemporal transport and degradation of pollutants. It ingests multi-source inputs—including environmental sensor streams and remote sensing data—to predict contaminant concentration fields in real time. To support practical decision-making, AquaDynNet is embedded into FlowSentinel, a management framework that enables real-time feedback, adaptive pollution control, and stakeholder-guided intervention planning. FlowSentinel integrates sensor updates through Bayesian correction and guides environmental policy via a participatory decision module. Figure 1 provides a high-level illustration of the proposed approach, showing how data flows through each component of the system, from collection to prediction and response. This modular design allows for accurate detection, dynamic adjustment, and transparent water quality governance.

This section outlines the mathematical formulation of water contamination as a dynamic system involving the transport and transformation of pollutants in a water body over time. The water body is represented as a spatial domain in three dimensions, where pollutants may enter through various sources such as agricultural runoff, wastewater discharge, or industrial effluents. The concentration of pollutants is treated as a function that depends on both space and time.

The rate at which pollutants enter the system is modeled using a source term that captures the input from different types of sources. These sources can be spatially distributed (agricultural runoff) or localized (industrial point discharge). The movement and spread of pollutants in water are governed by advection (transport by water flow) and diffusion (spreading due to mixing and turbulence).

To fully define the system, boundary conditions must be applied. These describe how contaminants behave at the boundaries of the water body. For example, Dirichlet conditions might specify a known concentration at a surface or riverbank, while Neumann conditions could represent scenarios where there is no pollutant flux across the boundary.

An important quantity to track is the total mass of contaminants in the system, which is obtained by integrating the concentration over the entire domain. Additionally, regulatory thresholds or environmental standards are defined for different contaminants. If the pollutant concentration exceeds these limits, the water quality is considered hazardous to human health and aquatic ecosystems.

In essence, the problem of monitoring and mitigating water contamination involves understanding the interplay between pollutant sources, water flow dynamics, diffusion processes, and spatial-temporal distribution of pollutants. By solving the underlying transport model, we can identify areas of high contamination and assess the long-term environmental impacts.

We introduce AquaDynNet, a novel model developed to simulate and predict water contamination dynamics with improved accuracy and environmental fidelity. This model integrates pollutant interactions, spatial heterogeneity, and adaptive flow dynamics to capture the complex behavior of contaminants in natural and engineered aquatic systems (As shown in Figure 2).

This section presents an advanced formulation of contaminant transport in aquatic environments by extending the classical advection-diffusion model. It incorporates key processes such as pollutant degradation, spatial heterogeneity in hydrodynamic and physicochemical parameters, and time-varying flow dynamics. These enhancements enable the model to better reflect real-world conditions such as biological and chemical decay, temperature effects, and turbulent dispersion. The transport dynamics also consider directionally variable mixing by using a tensor-based diffusion formulation, which is especially important in stratified or complex flow systems. The pollutant source term is modeled with stochastic elements to account for unpredictable events and spatial variability, including both continuous pollution and sudden releases. Additionally, the degradation rate of pollutants is modeled as a function of environmental conditions such as temperature, dissolved oxygen, and solar irradiance. This makes the model responsive to seasonal and diurnal fluctuations in the aquatic environment. To numerically solve the resulting equations, advanced computational techniques like finite-volume or spectral element methods are suggested, with mesh refinement in regions of steep gradients or variable pollution sources.

To capture the nuanced behavior of pollutant emissions, a spatiotemporal encoding of sources is introduced. This framework represents emission sources as spatially discrete and temporally dynamic, ideal for modeling scenarios like industrial discharges or stormwater outflows. Spatial singularities are first described using delta functions and then smoothed using Gaussian kernels for better numerical stability and accuracy. Emission intensity is modeled with both deterministic and stochastic elements, driven by environmental factors such as rainfall, temperature, and human activity schedules. This flexible design allows simulation of various pollution patterns including urban runoff, agricultural discharge, and industrial accidents. The framework also includes a method for calculating total pollutant load over a region and time window, which supports environmental compliance assessments and helps quantify the impact of pollution events.

To measure the changing impact of pollution on water quality, a time-dependent Water Quality Index (WQI) framework is introduced. This index integrates pollutant concentration, ecosystem sensitivity, and spatial regulatory thresholds to provide a comprehensive metric for ecological stress. It accounts for time-varying ecological sensitivity—such as seasonal events or vulnerable periods—and normalizes pollutant exposure against regulatory standards (As shown in Figure 3). For broader spatial evaluation, the WQI can be aggregated over a region using weights that reflect ecological or societal importance. The model also describes how the WQI evolves over time, allowing for analysis of both pollution events and mitigation efforts. The rate of change of WQI offers insights into whether water quality is improving or deteriorating in real time. By linking the WQI to predictive outputs from models such as AquaDynNet, the framework supports scenario-based forecasting and impact analysis. This enables integration into decision-support systems for both planning and response, offering a robust, adaptive, and predictive approach to water quality management. The WQI’s temporal and spatial responsiveness makes it a valuable tool for evaluating degradation and formulating mitigation strategies under dynamic environmental conditions.

In this section, we propose FlowSentinel, a comprehensive strategy for managing water contamination. Built upon the AquaDynNet model, FlowSentinel synergizes predictive simulation, real-time feedback, and participatory decision-making to create an intelligent, adaptive management framework for water quality control. The following three components form the strategic backbone of FlowSentinel (As shown in Figure 4).

FlowSentinel initiates its predictive feedback loop by integrating dense real-time sensor networks across aquatic systems. These sensors collect high-frequency data on various environmental variables such as pollutant concentration, water velocity, temperature, turbidity, dissolved oxygen, and precipitation. The data is continuously sent to a central processing unit, where the AquaDynNet model uses it to refine short-term forecasts. This predictive process functions as a recursive update system, where future pollutant concentrations are estimated based on current states and sensor inputs. The model not only updates the contaminant concentration but also adjusts parameters like flow velocity and source intensity. A Bayesian correction mechanism is incorporated to manage observational uncertainties, allowing the model to weigh possible outcomes based on how well they align with observed data. Furthermore, a discrepancy index is introduced to compare model predictions with real-time sensor observations, serving as a trigger for anomaly detection and alerts. This allows for the rapid identification of contamination events and enables prompt corrective actions. As a result, AquaDynNet transitions from a static model into an adaptive forecasting tool that evolves in response to real-world environmental changes and unexpected pollution episodes.

FlowSentinel utilizes an adaptive control architecture to optimize mitigation strategies under evolving environmental conditions. Real-time observations are mapped to control actions such as discharge limitations, flow regulations, land use adjustments, and ecological buffers. The model evaluates the effectiveness of intervention strategies by comparing predicted pollutant concentrations with regulatory or ecological targets. An optimization framework seeks to minimize deviation from these targets while accounting for implementation feasibility and cost. The system anticipates future pollutant dynamics using a rolling-horizon strategy, continuously updating as new sensor data becomes available. It also incorporates uncertainty through ensemble simulations, which explore various scenarios by perturbing input factors like flow rates, emission levels, and weather conditions. Ultimately, the optimized control strategies are translated into localized operational directives, ensuring that each subregion receives tailored guidance. This results in a spatially and temporally adaptive plan to maintain pollutant levels within acceptable limits, even in complex and variable aquatic environments.

To enable collaborative governance, FlowSentinel includes a participatory decision framework that integrates model-driven outputs with stakeholder-specific values and preferences. This system features a dynamic decision support interface for visualizing data, evaluating scenarios, and encoding stakeholder priorities. Stakeholders are modeled with distinct sensitivities to pollution impacts over space and time, allowing for personalized assessments of environmental exposure. The framework quantifies cumulative impacts and uses these metrics to compare different intervention scenarios. Preferences are expressed through multi-criteria evaluations balancing environmental benefits, economic costs, and equity considerations (As shown in Figure 5). The system supports trade-off analysis using Pareto frontiers and calculates impact disparities among stakeholders to identify strategies that reduce inequity. Stakeholder influence is spatially encoded, and collective decisions are made by maximizing a global utility function that balances stakeholder priorities across the region. Through this structure, FlowSentinel supports transparent, inclusive, and data-driven environmental management, integrating scientific modeling with stakeholder engagement for effective real-time governance.

This study leverages four complementary datasets that together offer a multi-dimensional view of freshwater pollution dynamics across space, time, and ecological impact. The first is the Terra Satellite Dataset (Venkata Sudhakar and Reddy., 2023), which provides high-resolution remote sensing imagery and surface reflectance data collected by NASA’s Terra spacecraft through its MODIS (Moderate Resolution Imaging Spectroradiometer) instrument. This dataset captures a range of geophysical variables relevant to water pollution modeling, including land surface temperature, vegetation indices, soil moisture anomalies, and surface runoff patterns. These variables are essential for contextualizing watershed-scale pollutant mobility and identifying anthropogenic influence zones, particularly in rural and peri-urban regions where ground-based monitoring is sparse. Supplementing this macro-scale view is the Aquatic Toxicity Dataset (Gajewicz-Skretna et al., 2021,), a curated collection of laboratory-tested toxicological responses of aquatic organisms to various chemical substances. It contains dose-response profiles for fish, invertebrates, and algae exposed to contaminants such as heavy metals, pesticides, and industrial solvents. The dataset is derived from standardized bioassays and peer-reviewed ecotoxicology studies, offering a quantitative foundation for linking contaminant concentrations to biological impact thresholds. This allows the model to not only track contaminant presence but also infer potential ecological damage through risk-weighted scoring. In addition to these remote and biological sources, we incorporate the Water Quality Dataset (Huang et al., 2021), a heterogeneous compilation of in situ water quality measurements collected from monitoring stations maintained by environmental agencies, hydrological networks, and public health institutions. Parameters measured include concentrations of nitrate, phosphate, lead, mercury, biological oxygen demand (BOD), chemical oxygen demand (COD), pH, turbidity, and dissolved oxygen (DO), all of which are instrumental in capturing the chemical state of the water body under observation. Temporal granularity varies from hourly to monthly depending on the station and region, enabling both short-term anomaly detection and long-term trend analysis. Geographical coverage spans multiple continents and climates, allowing for cross-regional model generalization and climate-resilience evaluation. Lastly, the WaterNet Dataset (Ajayi et al., 2022) provides a graph-structured representation of water distribution networks, drainage systems, and hydraulic connectivity within catchments and urban environments. Each node in the WaterNet graph corresponds to a hydrologic junction, such as a monitoring station, discharge outlet, or confluence point, while edges denote flow pathways, enabling the construction of a topologically consistent flow map. Attributes assigned to each node include flow rate, inflow composition, and contamination history, while edge weights represent hydraulic resistance, transport delay, or dispersion coefficients. WaterNet serves as the computational backbone for simulating contaminant propagation through infrastructure-dense systems, particularly under conditions of rapid flow changes such as stormwater events or infrastructure failures. When integrated, these four datasets support a comprehensive and hierarchical modeling strategy, from global land-surface predictors and ecological hazard characterization to localized chemical readings and network-structured flow dynamics. This multi-source integration enhances the robustness of pollutant forecasts and supports scalable application across geographies with differing data infrastructure maturity. Furthermore, the diverse data modalities foster opportunities for multimodal learning, enabling the model to cross-validate predictions from independent data views and to fill in missing information through representational transfer. Together, these datasets form the empirical foundation of our system, guiding not only the calibration and validation of simulation modules but also the evaluation of mitigation strategies within the FlowSentinel framework.

In order to enhance transparency and reproducibility, we now provide detailed characteristics of the four datasets used. The Terra Satellite Dataset consists of high-resolution remote sensing imagery collected via NASA’s MODIS instrument onboard the Terra spacecraft. It features a spatial resolution of 500 m and a temporal frequency ranging from 1 to 8 days, depending on the variable. The dataset includes over 20,000 image tiles across diverse geographic regions, with global coverage spanning all continents, which supports generalized training for pollution pattern recognition. The Aquatic Toxicity Dataset comprises approximately 5,100 curated entries detailing toxicological responses of aquatic organisms—fish, invertebrates, and algae—under controlled laboratory conditions. Each entry includes dosage-response relationships for one contaminant-organism pair, and the dataset spans chemical exposures from over 30 countries across Asia, Europe, and the Americas. Temporal metadata indicates experimental collection primarily between 2005 and 2021. The Water Quality Dataset contains over 12,000 field-recorded entries from environmental monitoring stations. Measurements include nitrate, phosphate, pH, turbidity, and dissolved oxygen, with sampling frequencies ranging from hourly to monthly. The spatial scope encompasses rivers and lakes across 18 countries, including extensive coverage in China, Germany, the United States, and India. The dataset supports both real-time anomaly detection and long-term trend analysis. The WaterNet Dataset models hydraulic infrastructure via a graph representation. It contains over 2,000 nodes and 3,500 edges representing hydrological junctions and flow pathways in urban, peri-urban, and rural environments. It covers networks in 15 countries with a mix of topologies and includes flow rate, inflow composition, and historical contamination data. Temporal dynamics are encoded through hourly simulations and logged events, enabling time-aware modeling of pollutant propagation through infrastructure systems.

To assess the effectiveness of the proposed framework, we conduct extensive experiments using four representative benchmark datasets: Terra Satellite, Aquatic Toxicity, Water Quality, and WaterNet. Each dataset is divided into stratified training and testing subsets to maintain balanced distributions across spatial, temporal, and ecological dimensions. Performance outcomes are benchmarked against leading existing methods under identical conditions to guarantee experimental fairness and replicability. Our unified model architecture features a convolutional neural network (CNN) backbone enriched with spatial attention layers, followed by multiple fully connected modules. The CNN component is responsible for capturing multi-resolution features from multimodal inputs, while the attention mechanism directs the model’s focus toward spatial-temporal regions most indicative of pollution signals. Prior to training, all input data are carefully preprocessed. For spatial data such as satellite imagery, operations include normalization, per-channel standardization, and image patch extraction. For graph-based datasets like WaterNet, we apply graph-level normalization along with spectral embedding techniques to preserve relational structures. We further augment the data through a variety of techniques, including geometric transformations (e.g., rotations, scalings, flips), elastic deformations, brightness modulation, and temporal jittering to enhance model robustness and reduce overfitting. Training is carried out using the Adam optimizer with an initial learning rate of 0.001 and a batch size set to 32. The training process spans 100 epochs, with an early stopping criterion triggered if no improvement is observed in the validation loss over 10 successive epochs. All experiments are executed on an NVIDIA GPU (Model X) equipped with 16 GB of memory to support high-throughput processing of large-scale multimodal data. Regularization methods include L2 penalty (with a decay factor of 0.0001) and dropout (set at 0.5) applied to the dense layers to mitigate overfitting. We evaluate model performance using standard classification metrics such as accuracy, precision, recall, and F1-score. For segmentation-related tasks, we report the mean Intersection-over-Union (mIoU). For datasets exhibiting class imbalance, especially in toxicity prediction tasks, the area under the receiver operating characteristic curve (AUC) is also included. Each experiment is conducted independently three times with varying random seeds, and the final performance is reported as the mean along with standard deviation to capture variability. Additionally, we perform ablation experiments to isolate the contribution of each component (e.g., graph integration, multimodal inputs), and apply statistical significance testing to determine the reliability of observed improvements over competitive baselines.

On the Terra Satellite dataset, our method outperforms all other methods, achieving a notable accuracy of 90.75 ± 0.02, recall of 86.41 ± 0.02, F1 score of 88.79 ± 0.01, and AUC of 92.02 ± 0.03. Compared to the best-performing method (BLIP), our method shows a significant improvement, especially in recall and F1 score, which indicates better generalization and robustness. Similarly, on the Aquatic Toxicity dataset, AquaDynNet achieves an accuracy of 92.58 ± 0.02, recall of 89.73 ± 0.02, F1 score of 91.12 ± 0.02, and AUC of 94.13 ± 0.03, which again surpasses all other methods, including BLIP, ViT, and I3D. On the Water Quality dataset, AquaDynNet achieves an accuracy of 88.34 ± 0.02, recall of 83.68 ± 0.03, F1 score of 85.54 ± 0.02, and AUC of 89.72 ± 0.03, outperforming the best alternative method (BLIP). On WaterNet, AquaDynNet achieves an accuracy of 91.98 ± 0.03, recall of 87.61 ± 0.02, F1 score of 89.22 ± 0.03, and AUC of 92.47 ± 0.02, further demonstrating its superior performance across all datasets. The significant improvements in performance achieved by AquaDynNet can be attributed to its novel architecture and the effective integration of different features from various data sources. Our method effectively captures the complex relationships within the datasets, leading to higher accuracy, recall, F1 score, and AUC, especially in comparison to existing methods.

The results highlight the effectiveness and superiority of AquaDynNet over current SOTA methods, making it a promising approach for 3D object recognition and related tasks.

The experimental results in Table 1; Table 2 consistently demonstrate that AquaDynNet outperforms competing models across all benchmark datasets. Notably, the model achieves a significant AUC improvement over ViT and BLIP, two strong baselines in vision-language and spatiotemporal modeling. This improvement is attributed to the incorporation of physically-informed components such as the Generalized Transport Dynamics and Spatiotemporal Source Encoding modules, which enable the model to simulate real-world pollutant behavior more effectively. The model’s superior recall and F1-scores suggest it is particularly adept at detecting contamination events without missing rare or spatially dispersed pollution signals—a critical requirement for environmental monitoring. The high AUC on the WaterNet dataset underscores its robustness in complex hydraulic infrastructures, where pollutant dispersion is affected by dynamic network topology. Furthermore, the low variance in performance across three independent runs highlights the model’s stability and generalization. These observations collectively affirm that our approach is not only accurate but also interpretable and resilient across varying water quality challenges.

The models used for comparison—CLIP, ViT, I3D, BLIP, Wav2Vec 2.0, and T5—are not originally designed for environmental modeling. However, they represent cutting-edge architectures that are widely adopted for tasks involving multimodal learning, satellite imagery analysis, and temporal sequence modeling. These characteristics align with the core demands of our problem setting, where input data consists of remote sensing images, spatially distributed sensor streams, and time-series measurements of water quality parameters. ViT and I3D are capable of processing temporal sequences and learning from spatiotemporal data, making them suitable for modeling contamination trends over time. CLIP and BLIP, although from the vision-language domain, offer robust image representation capabilities that can generalize to remote sensing contexts due to their pretraining on large and diverse datasets. Their adaptability to various data modalities makes them strong benchmarks for evaluating the generalizability of AquaDynNet. Wav2Vec 2.0 and T5 were included for their sequence modeling strength and ability to process structured inputs such as environmental logs or time-tagged pollutant records. Conventional domain-specific baselines—such as hydrodynamic simulation tools (SWAT, HEC-RAS) or rule-based environmental models—were not included due to fundamental architectural differences. These models are not designed for end-to-end learning with multimodal inputs and typically lack the capability for real-time prediction or integration with sensor feedback. Many rely heavily on manual parameter tuning and domain-specific calibration, which limits their scalability and reproducibility in AI-driven pipelines. We have thus focused on models that align better with the data modalities and operational goals of our framework.

To evaluate the individual impact of each fundamental component within our AquaDynNet framework, we perform a comprehensive ablation analysis by sequentially excluding three essential modules. Generalized Transport Dynamics, Spatiotemporal Source Encoding, and Real-Time Sensing Integration. The results on the Terra Satellite and Aquatic Toxicity datasets are reported in Tables 3, 4. We observe that removing the Generalized Transport Dynamics module results in the most significant degradation, especially in accuracy and AUC, underscoring the importance of physically grounded contaminant propagation modeling. On the Terra dataset, for instance, accuracy drops from 90.75% to 85.34%, and AUC from 92.02 to 85.90, illustrating that spatial-temporal diffusion effects are essential to model contaminant flow patterns driven by hydrological transport. When the Spatiotemporal Source Encoding module is removed, recall and F1 score decline across both datasets, demonstrating the importance of capturing dynamic, localized pollutant inputs, such as episodic discharges or spatially clustered contamination events. The removal of Real-Time Sensing Integration also leads to noticeable performance reduction, albeit to a lesser extent, suggesting that online observation correction provides added robustness but is particularly beneficial in settings with dynamic environmental inputs.

The ablation results are presented on the Water Quality and WaterNet datasets. The performance trends observed here are consistent with those from the previous datasets, but the importance of Real-Time Sensing Integration becomes more pronounced, especially on the WaterNet dataset. Removing this component leads to a substantial drop in all metrics, with F1 score decreasing from 89.22 to 84.05 and AUC from 92.47 to 88.41, indicating the critical role of real-time environmental updates in infrastructure-dense and highly dynamic water systems. In contrast, the removal of Generalized Transport Dynamics yields large accuracy and AUC penalties on Water Quality data, confirming the necessity of modeling pollutant advection and degradation under varying hydrological and chemical conditions. The ablation of Spatiotemporal Source Encoding consistently results in lower recall and F1 scores, reflecting a reduced ability to detect distributed or time-variant contamination sources. These findings demonstrate that each component within AquaDynNet plays a distinct and irreplaceable role across different data modalities and environmental settings, and validate the model’s composite architecture as essential for generalizable performance in real-world water quality modeling tasks.

The ablation study presented in Table 3; Table 4 provides critical insights into the contribution of each core module in AquaDynNet. Removing the Generalized Transport Dynamics (GTD) module leads to the largest degradation in both accuracy and AUC across all datasets, confirming that embedding hydrodynamic priors into the model is essential for capturing realistic pollutant behavior. Specifically, AUC dropped by 6.12 points on the Water Quality dataset and 6.46 points on WaterNet when GTD was removed, underscoring its role in preserving physical consistency. Excluding the Spatiotemporal Source Encoding (STE) led to decreased recall and F1-score, suggesting a reduced ability to handle temporally variable or spatially concentrated emissions. This module is particularly valuable in scenarios involving episodic discharges or spatial anomalies, such as industrial runoff or urban stormwater events. The Real-Time Sensing Integration (RTSI) module, when removed, caused notable performance drops in WaterNet, where sensor updates are most critical due to the rapidly changing environmental conditions of infrastructure-heavy systems. The AUC drop from 92.47 to 88.41 further highlights its role in dynamic adaptation. Overall, these experiments validate the complementary strengths of each architectural component and justify the integrated design of AquaDynNet for generalizable and interpretable water contamination detection.

To provide a clearer understanding of how each architectural component contributes to the overall performance of AquaDynNet, we present a comparative visualization in Figure 6. This line plot captures the metric changes—Accuracy, F1-Score, and AUC—across four datasets (Terra, Aquatic Toxicity, Water Quality, and WaterNet) under different ablation settings. The Full configuration represents the complete model, while the subsequent points indicate performance when one core component is removed: Generalized Transport Dynamics (G.T.D.), Spatiotemporal Encoding (S.T.E.), or Real-Time Sensing Integration (R.T.S.I.). The visualization reveals that the absence of G.T.D. leads to the sharpest performance declines across all metrics, particularly on the Terra and Water Quality datasets, underscoring its central role in modeling physical transport and pollutant dynamics. Similarly, excluding R.T.S.I. results in significant performance drops, most notably in WaterNet, where real-time feedback is essential for handling rapidly changing conditions. Although removing S.T.E. also impacts performance, the decline is comparatively moderate. By providing a consolidated view across all datasets and metrics, this figure confirms that each module within AquaDynNet plays a unique and essential role. The composite design ensures the model’s high accuracy, robustness, and applicability in diverse ecological settings. This visual summary strengthens our earlier quantitative findings and offers an intuitive complement to the ablation tables.

To clarify the modeling objectives and comparative structure of our experiments, we organize this section into three specific layers: task definition, baseline comparison, and performance analysis. For each dataset, the modeling goal is clearly defined. These comparisons are summarized in Table 5, highlighting how AquaDynNet outperforms state-of-the-art baselines and demonstrating the critical role of each architectural module. For instance, in the Water Quality dataset, the objective is to classify water samples into safe or contaminated categories based on sensor measurements. In contrast, the WaterNet dataset emphasizes contaminant propagation modeling across hydraulic networks, requiring both node-level prediction and temporal trend forecasting. We compared our model (AquaDynNet) against six representative baseline methods: CLIP, ViT, I3D, BLIP, Wav2Vec 2.0, and T5. Each method was chosen to reflect either a strong visual-language foundation or a temporal sequence modeling capability. All models were trained and tested under identical conditions using stratified dataset splits to ensure fairness. Evaluation metrics include accuracy, recall, F1-score, and AUC. These metrics were calculated across three independent trials, with mean and standard deviation reported. To make the comparison more interpretable, we added specific examples to the text. For instance, on the Water Quality dataset, AquaDynNet achieved an F1-score of 85.54 ± 0.02, compared to 82.17 ± 0.01 by the best-performing baseline BLIP, reflecting a 3.37% absolute improvement. On the Terra Satellite dataset, AquaDynNet’s AUC score reached 92.02 ± 0.03, substantially higher than ViT’s 85.98 ± 0.02. These differences demonstrate the robustness and generalizability of AquaDynNet across modalities and tasks. In the ablation study, we tested the effect of removing each core module in isolation. When the Generalized Transport Dynamics module was excluded, the accuracy on the Aquatic Toxicity dataset dropped from 92.58% to 88.11%, with the AUC decreasing from 94.13 to 89.45. This confirms the value of physically grounded pollutant dynamics modeling. Similarly, the absence of Real-Time Sensing Integration caused a significant drop in performance, particularly in dynamic datasets like WaterNet, suggesting that our real-time feedback mechanism is crucial for accuracy and responsiveness. This restructuring and the added examples aim to make the modeling pipeline and performance implications more transparent and verifiable for readers.

While the above experiments focus on evaluating model performance across diverse datasets, our proposed framework is explicitly designed for real-world deployment in water quality monitoring systems. The modular architecture of AquaDynNet and FlowSentinel supports seamless integration with real-time environmental sensing platforms. For example, AquaDynNet accepts input from distributed water quality sensors deployed in rivers, lakes, or reservoirs, which can continuously stream measurements such as pH, turbidity, temperature, and dissolved oxygen. These data are directly assimilated into the model pipeline through the FlowSentinel component. FlowSentinel functions as a dynamic control loop that receives real-time observational data and adjusts AquaDynNet’s predictions using Bayesian correction and Kalman filtering. This allows the system to account for unexpected pollution events, missing sensor values, or external environmental shocks. The system can be embedded within existing hydrological infrastructure through edge computing units that collect, preprocess, and relay sensor signals to the central prediction engine. Our framework also supports participatory decision-making by allowing stakeholders—including policymakers, environmental agencies, and utility operators—to interact with model outputs through a web-based interface. They can simulate hypothetical pollution scenarios, evaluate mitigation strategies, and prioritize intervention locations based on the predicted water quality index and ecological risk estimates. Since the model is grounded in physical transport dynamics, it provides not only accurate forecasts but also interpretable spatial patterns that align with how contamination disperses in real water bodies. This operational setup transforms AquaDynNet from a static prediction model into an adaptive, closed-loop monitoring system suitable for continuous deployment. It enables predictive surveillance and proactive intervention, making it well-suited for managing water quality in complex ecological regions and under resource-constrained monitoring conditions.