TLT: Recurrent fine-tuning transfer learning for water quality long-term prediction

Interpretable prediction, classification and regulation of water quality: A case study of Poyang Lake, China

Effective identification and regulation of water quality impact factors is essential for water resource management and environmental protection. However, the complex coupling of water quality systems poses a significant challenge to this task. This study proposes coherent model for water quality prediction, classification and regulation based on interpretable machine learning. The decomposition-reconstruction module is used to transform non-stationary water quality series into stationary series while effectively reducing the feature dimensions. Spatiotemporal multi-source data is introduced by using the Maximum Information Coefficient (MIC) for feature selection. The Temporal Convolutional Network (TCN) is used to extract the temporal features of different variables, followed by the introduction of External Attention mechanism (EA) to construct the relationship between these features. Finally, the target water quality sequence is simulated using Gated Recurrent Unit (GRU). The proposed model was applied to Poyang Lake in China to predict six water quality indicators: ammonia nitrogen (NH3-N), dissolved oxygen (DO), pH, total nitrogen (TN), total phosphorus (TP), water temperature (WT). The water quality was then classified based on the prediction results using the XGBoost algorithm. The findings indicate that the proposed model's Nash-Sutcliff Efficiency (NSE) value ranges from 0.88 to 0.99, surpassing that of the benchmark model, and demonstrates strong interval prediction performance. The results highlight the superior performance of the XGBoost algorithm (with an accuracy of 0.89) in addressing water quality classification issues, particularly in cases of category imbalance. Subsequently, interpretability analysis using the SHapley Additive exPlanation (SHAP) method revealed that the model is capable of learning relationships between different variables and there exists a possibility of learning the physical laws. Ultimately, this study proposes a water quality regulation mechanism that improves TN and DO levels by stepwise changing the magnitude of water temperature, which significantly improves in the case of data limitations. In conclusion, this study presents an overall framework for integrating water quality prediction, classification and improvement for the first time, forming a complete set of water quality early warning and improvement management strategies. This framework provides new ideas and ways for lake water quality management.

Which riverine water quality parameters can be predicted by meteorologically-driven deep learning?

River water quality has been significantly impacted by climate change and extreme weather events worldwide. Despite increasing studies on deep learning techniques for river water quality management, understanding which riverine water quality parameters can be well predicted by meteorologically-driven deep learning still requires further investigation. Here we explored the prediction performance of a traditional Recurrent Neural Network, a Long Short-Term Memory network (LSTM), and a Gated Recurrent Unit (GRU) using meteorological conditions as inputs in the Dahei River basin. We found that deep learning models (i.e., LSTM and GRU) demonstrated remarkable effectiveness in predicting multiple water quality parameters at daily scale, including water temperature, dissolved oxygen, electrical conductivity, chemical oxygen demand, ammonia nitrogen, total phosphorous, and total nitrogen, but not turbidity. The GRU model performed best with an average determination coefficient of 0.94. Compared to the daily-average prediction, the GRU model exhibited limited error increment of 10-40 % for most water quality parameters when predicting daily extreme values (i.e., the maximum and minimum). Moreover, deep learning showed superior performance in collective prediction for multiple water quality parameters than individual ones, enabling a more comprehensive understanding of the river water quality dynamics from meteorological data. This study holds the promise of applying meteorologically-driven deep learning techniques for water quality prediction to a broader range of watersheds, particularly in chemically ungauged areas.

Enhanced forecasting of chlorophyll-a concentration in coastal waters through integration of Fourier analysis and Transformer networks

The accurate prediction of chlorophyll-a (chl-a) concentration in coastal waters is essential to coastal economies and ecosystems as it serves as the key indicator of harmful algal blooms. Although powerful machine learning methods have made strides in forecasting chl-a concentrations, there remains a gap in effectively modeling the dynamic temporal patterns and dealing with data noise and unreliability. To wiggle out of quagmires, we introduce an innovative deep learning prediction model (termed ChloroFormer) by integrating Transformer networks with Fourier analysis within a decomposition architecture, utilizing coastal in-situ data from two distinct study areas. Our proposed model exhibits superior capabilities in capturing both short-term and middle-term dependency patterns in chl-a concentrations, surpassing the performance of six other deep learning models in multistep-ahead predictive accuracy. Particularly in scenarios involving extreme and frequent blooms, our proposed model shows exceptional predictive performance, especially in accurately forecasting peak chl-a concentrations. Further validation through Kolmogorov-Smirnov tests attests that our model not only replicates the actual dynamics of chl-a concentrations but also preserves the distribution of observation data, showcasing its robustness and reliability. The presented deep learning model addresses the critical need for accurate prediction on chl-a concentrations, facilitating the exploration of marine observations with complex dynamic temporal patterns, thereby supporting marine conservation and policy-making in coastal areas.

Integration of interpretable machine learning and environmental magnetism elucidates reduction mechanism of bioavailable potentially toxic elements in lakes after monsoon

Little information is available on the influence of substantial precipitation and particulate matter entering during the monsoon process on the release of potentially toxic elements (PTEs) into lake sediments. Sediments from a typical subtropical lake across three periods, pre-monsoon, monsoon, and post-monsoon, were collected to determine the chemical forms of 12 PTEs (As, Cd, Co, Cr, Cu, Fe, Hg, Pb, Mn, Ni, Sb, and Zn), magnetic properties, and physicochemical indicators. Feature importance, Shapley additive explanations, and partial dependence plots were used to explore the factors influencing bioavailable PTEs. The proportion of bioavailable forms of PTEs decreased from 3.85 % (Cd) to 87.84 % (Hg) after the monsoon. Gradient extreme boosting demonstrated robust fitting accuracy for the prediction of the bioavailable forms of the 12 PTEs (R2 > 0.84). Shapley additive explanations identified that the bioavailable forms were influenced by the total PTE concentrations, wind, shortwave radiation, and particle inputs (25.1 %-88.5 % for total importance), either individually or in combination. The partial dependence plots highlighted the influence thresholds of background values and anthropogenic factors on the bioavailable forms of PTEs. Changes in environmental properties could indicate the process of external sediment influx into lakes. The optimized model combined with magnetic parameters showed strong performance in other cases (coefficient of determination>0.58), confirming the ubiquitous decrease in bioavailable forms of PTEs in sediments across subtropical lakes after monsoons.

Stream salinity prediction in data-scarce regions: Application of transfer learning and uncertainty quantification

Scarcity of stream salinity data poses a challenge to understanding salinity dynamics and its implications for water supply management in water-scarce salt-prone regions around the world. This paper introduces a framework for generating continuous daily stream salinity estimates using instance-based transfer learning (TL) and assessing the reliability of the synthetic salinity data through uncertainty quantification via prediction intervals (PIs). The framework was developed using two temporally distinct specific conductance (SC) datasets from the Upper Red River Basin (URRB) located in southwestern Oklahoma and Texas Panhandle, United States. The instance-based TL approach was implemented by calibrating Feedforward Neural Networks (FFNNs) on a source SC dataset of around 1200 instantaneous grab samples collected by United States Geological Survey (USGS) from 1959 to 1993. The trained FFNNs were subsequently tested on a target dataset (1998-present) of 220 instantaneous grab samples collected by the Oklahoma Water Resources Board (OWRB). The framework's generalizability was assessed in the data-rich Bird Creek watershed in Oklahoma by manipulating continuous SC data to simulate data-scarce conditions for training the models and using the complete Bird Creek dataset for model evaluation. The Lower Upper Bound Estimation (LUBE) method was used with FFNNs to estimate PIs for uncertainty quantification. Autoregressive SC prediction methods via FFNN were found to be reliable with Nash Sutcliffe Efficiency (NSE) values of 0.65 and 0.45 on in-sample and out-of-sample test data, respectively. The same modeling scenario resulted in an NSE of 0.54 for the Bird Creek data using a similar missing data ratio, whereas a higher ratio of observed data increased the accuracy (NSE = 0.84). The relatively narrow estimated PIs for the North Fork Red River in the URRB indicated satisfactory stream salinity predictions, showing an average width equivalent to 25 % of the observed range and a confidence level of 70 %.

Innovative interpretable AI-guided water quality evaluation with risk adversarial analysis in river streams considering spatial-temporal effects

Water security remains a critical issue given the looming threats of industrial pollution, necessitating comprehensive assessments of water quality to address seasonal fluctuations and influential factors while formulating effective strategies for decision makers. This study introduces a novel approach for evaluating water quality within a complex riverine zone in South Korea: Han River that encompasses five river streams situated at each junction of North and South streams (including Gyeongan Stream) that ultimately leading towards Paldang Lake. By utilizing the monthly water characteristic data from the year 2013-2022 across 14 different locations, the significant seasonal trends and potential influences on water quality are identified. The water quality here is calculated with the proposed method of sub-index water quality index (s-WQI). A combinatorial prediction approach of s-WQI for each location is conducted through a collective of data preprocessing approaches including Hampel filtering and feature selection in prior to the machine learning predictions. In return, light gradient boosting (LGB) is the most accurate predictor by outperforming other prediction algorithms, especially through LGB-Pearson and LGB-Spearman combinations for North and South stream intersections, and LGB-Pearson for Paldang Lake. To further evaluate the robustness of this evaluation and extending the results to a foreseeable scenario, a seasonal based Monte-Carlo Simulation with 10,000 attempts targeting the water characteristic distributions obtained from each location considered are carried out to identify the risk bounds within. The results are further interpreted with SHAP analysis on identifying the contributions of each water characteristics towards the water quality through local and global spectrum. This research yields practical implications, offering tailored strategies for water quality enhancement and early warning systems. The integration of AI-based prediction and feature selection underscores the transformative potential of computational techniques in advancing data-driven water quality assessments, shaping the future of environmental science research.

Deep learning in water protection of resources, environment, and ecology: achievement and challenges

The breathtaking economic development put a heavy toll on ecology, especially on water pollution. Efficient water resource management has a long-term influence on the sustainable development of the economy and society. Economic development and ecology preservation are tangled together, and the growth of one is not possible without the other. Deep learning (DL) is ubiquitous in autonomous driving, medical imaging, speech recognition, etc. The spectacular success of deep learning comes from its power of richer representation of data. In view of the bright prospects of DL, this review comprehensively focuses on the development of DL applications in water resources management, water environment protection, and water ecology. First, the concept and modeling steps of DL are briefly introduced, including data preparation, algorithm selection, and model evaluation. Finally, the advantages and disadvantages of commonly used algorithms are analyzed according to their structures and mechanisms, and recommendations on the selection of DL algorithms for different studies, as well as prospects for the application and development of DL in water science are proposed. This review provides references for solving a wider range of water-related problems and brings further insights into the intelligent development of water science.

A coupled model to improve river water quality prediction towards addressing non-stationarity and data limitation

Accurate predictions of river water quality are vital for sustainable water management. However, even the powerful deep learning model, i.e., long short-term memory (LSTM), has difficulty in accurately predicting water quality dynamics owing to the high non-stationarity and data limitation in a changing environment. To wiggle out of quagmires, wavelet analysis (WA) and transfer learning (TL) techniques were introduced in this study to assist LSTM modeling, termed WA-LSTM-TL. Total phosphorus, total nitrogen, ammonia nitrogen, and permanganate index were predicted in a 4 h step within 49 water quality monitoring sites in a coastal province of China. We selected suitable source domains for each target domain using an innovatively proposed regionalization approach that included 20 attributes to improve the prediction efficiency of WA-LSTM-TL. The coupled WA-LSTM facilitated capturing non-stationary patterns of water quality dynamics and improved the performance by 53 % during testing phase compared to conventional LSTM. The WA-LSTM-TL, aided by the knowledge of source domain, obtained a 17 % higher performance compared to locally trained WA-LSTM, and such improvement was more impressive when local data was limited (+66 %). The benefit of TL-based modeling diminished as data quantity increased; however, it outperformed locally direct modeling regardless of whether target domain data was limited or sufficient. This study demonstrates the reasoning for coupling WA and TL techniques with LSTM models and provides a newly coupled modeling approach for improving short-term prediction of river water quality from the perspectives of non-stationarity and data limitation.

Water quality analysis based on LSTM and BP optimization with a transfer learning model

In the urban water environmental management, a fast and effective method for water quality analysis should be established with the rapid urbanization. In this study, the Beijing's sub-center was chosen as a case study, and long short-term memory (LSTM) and back propagation (BP) models were built, then a transfer learning model was proposed and applied to optimize the two models on the base of the upstream and downstream relationships in the rivers. The results indicated that the proposed deep learning model could improve NSE by 7% and 9% for LSTM and BP at the Dongguan Bridge gauge, respectively. At the Xugezhuang gauge in the Liangshui River, NSE was improved by 11% and 17%, respectively. At the Yulinzhuang gauge, it was improved by 16% and 13%, respectively. Because the upstream and downstream relationships were considered in the learning model, the model performance was obviously better. In brief, this method would provide an idea for the effective water quality model construction in the ungauged basins or regions.

Multi-sensor and multi-platform retrieval of water chlorophyll a concentration in karst wetlands using transfer learning frameworks with ASD, UAV, and Planet CubeSate reflectance data

China has one of the widest distributions of carbonate rocks in the world. Karst wetland is a special and important ecosystem of carbonate rock regions. Chlorophyll-a (Chla) concentration is a key indicator of eutrophication, and could quantitatively evaluate water quality status of karst wetland. However, the spectral reflectance characteristics of the water bodies of karst wetland are not yet clear, resulting in remote sensing retrieval of Chla with great challenges. This study is a pioneer in utilizing field-based full-spectrum hyperspectral data to reveal the spectral response characteristics of karst wetland water body and determine the sensitive spectral bands of Chla. We further evaluated the Chla retrieval performance of multi-platform spectral data between Analytical Spectral Device (ASD), Unmanned aerial vehicle (UAV), and PlanetScope (Planet). We proposed two multi-sensor weighted integration strategies and two transfer learning frameworks for estimating water Chla from the largest karst wetland in China by combing a partial least square with adaptive ensemble algorithms. The results showed that: (1) In the range of 500-850 nm, the spectral reflectance of water bodies in the karst wetland was overall 0.001-0.105 higher than the inland water bodies, and the sensitive spectral ranges of water Chla focus on 603-778 nm; (2) UAV images outperformed ASD and Planet data, and produced the highest inversion accuracy (R2 = 0.670) for water Chla in karst wetland; (3) Multi-sensor weighted integration retrieval methods improved the Chla estimation accuracy (R2 = 0.716). Integration retrieval methods with the different weights produced the better Chla estimation accuracy than that of methods with the equal weights; (4) The transfer learning methods from ASD to UAV platform provided the better retrieval performance (the average R2 = 0.669) than that of methods from UAV to Planet platform. The transfer learning methods obtained the highest estimation accuracy of Chla (R2 = 0.814) when the ratio of the training and test data in the target domain was 7:3. The transfer learning methods produced the higher estimation accuracies with the distribution of the absolute residuals between predicted and measured values <20.957 mg/m3 compared to the multi-sensor weighted integration retrieval methods, which demonstrated that transfer learning is more suitable for estimating Chla in karst wetland water bodies using multi-platform and multi-sensor data. The results provide a scientific basis for the protection and sustainable development of karst wetlands.