Sensing_and_Reasoning_of_Water_Quality_Based_on_Deep_Reinforcement_Learning_in_Complex_Watershed.pdf
Abstract—Aquatic information monitoring is crucial for the sustainable management of water environments. Conventional interpolation methods commonly hinge on assumptions of spatial proximity or temporal similarity. However, they often fall short of capturing the intricate spatiotemporal correlations present in water quality sequences, affecting our understanding of the spatial patterns of regional water quality conditions. In this study, we propose a framework for river basin information fine-grained sensing based on deep learning, which includes a global sensing model and a static deployment model. Inside the global sensing model, we adopt a multi-dimensional convolutional neural network (CNN) to extract spatiotemporal features and an attention mechanism to fuse these features, to infer water quality variable information on unmonitored points. Since the inference outcomes could be affected by the locations of the sensors, to minimize the inference error of the global sensing model, the static deployment model was designed to aid the deployment of sensors into strategic locations of a river basin to obtain optimum spatial-temporal data samples. The research results not only revealed the spatial distribution patterns of total nitrogen (TN) concentrations but also showed that the proposed method could yield a better inference performance compared to traditional interpolation methods.
Index Terms—Water Quality, Sensing, Inference, Spatiotem- poral, CNN, Sensor Deployment
IN recent years, widespread water pollution has intensified the impact of water scarcity on the socio-economic fabric. Non-point source pollution, arising from agricultural production and animal husbandry operations, is a major cause of water quality damage. On one hand, non-point source pollution is difficult to detect and manage in a timely manner due to its diffusion properties. On the other hand, water quality monitoring programs aimed at river protection and management remain limited . Therefore, increasing the monitoring of aquatic environments is an urgent matter. Due to the extensive physical labor and the material costs required for deploying sensors or other measuring instruments, the collection and analysis of water samples are often expensive . The high cost of water sampling and analysis limits the number of routine water quality monitoring
stations. In China, national routine monitoring points are mainly concentrated along the main rivers and their major tributaries. For instance, the Huai River basin, which has a total drainage area of 270,000 square kilometers, has only 10 national routine monitoring points along its main course, and its major tributaries have only 42 monitoring points [7]. This is also the case in many other countries worldwide. Table I provides a statistic of the number of monitoring stations along major rivers in some European countries in 2023. TThe table shows that many countries have a low density of monitoring stations, especially in the regions of major rivers, where monitoring station deployment is limited. Additionally, the establishment of monitoring stations mainly focuses on the mainstream, leaving many tributaries without adequate coverage. This situation hinders accurate assessment of water quality and the timely detection of water quality issues.
Limited monitoring station coverage for watershed water quality can impact our understanding of regional water qual ity spatial patterns and hinder the development of effective watershed pollution control strategies. A viable solution is estimating water quality of unmonitored locations using exist ing observations, utilizing the relationships between different spatial water quality parameters and watershed characteristics. In our study, we explored the capabilities of data-driven deep learning techniques to construct effective regional models for water quality data. We proposed an integrated sensing solution for water quality monitoring based on deep learning and reinforcement learning. Based on non-uniform sensing nodes, this solution can generate deployment strategies for var ious sensing environments to optimize sensor placement and collect specific water quality parameters. The global sensing model is then applied for spatiotemporal feature extraction and fusion of multidimensional data, ultimately obtaining more comprehensive water quality monitoring data in a complex watershed. The contributions of this study are as follows: (1) We propose a global sensing model for inferring water quality variables, which can accommodate various forms of irregular samples as input. (2) A sensor deployment model based on deep reinforce ment was designed to facilitate efficient decision-making re garding the optimal sensor deployment locations in river basins so as to minimize inference errors of global sensing model. (3) As a result of the performance evaluation, it was demonstrated that our method can deliver higher inference accuracies compared to traditional interpolation methods in the Korean river water quality dataset.
In this paper, a methodological approach for sensing and inference of river water quality has been introduced and tested. GSM-SDMconsists of two parts: a global sensing model and a
sensors deployment model. The global sensing model extracts and fuses spatiotemporal features from sparsely sampled data to enable accurate inference of unobserved spatiotemporal points. The sensor deployment model minimizes estimation errors in information inference by finding key deployment points. Experiments were conducted on real-world water qual ity data collected from the Geum and Nakdong River basins in Korea, and the results indicate that our method not only accurately reveals the spatial patterns of TN distribution, but also has significant advantages over traditional methods. The proposed method can also be applied to measure other water quality variables for which sensors are expensive or difficult to maintain, including metals, nutrients, and algal biomass.