Advancing Water Quality Management: Harnessing the Synergy of Remote Sensing, Process-Based Models, and Machine Learning to Enhance Monitoring and Prediction

Abstract

Amid the intensifying challenges of climate change and human activities such as shifts in agricultural practices, the pressure on water resources, particularly regarding water quality, has intensified. As a result, improving water quality monitoring and prediction has emerged as an essential strategy to tackle these challenges and ensure the sustainable management of water resources. Traditional water quality monitoring technologies have inherent limitations; however, integrating remote sensing (RS) technologies with modeling approaches has shown significant promise in enhancing water quality monitoring and prediction. This integrated approach significantly improves the accuracy and intelligence of monitoring and prediction, while extending spatiotemporal coverage, lowering monitoring costs, and enabling more comprehensive analysis through optimized monitoring design, multi-source data fusion, and the synergistic coupling of data-driven and process-based models (PBMs). Advanced models, particularly those combining PBMs with AI techniques, further enhance predictive capabilities for water quality. Despite these advances, the application of these integrated methods faces challenges in areas such as data management, monitoring elusive pollutants, model accuracy and efficiency, system integration, and real-world implementation. In response to these challenges, this paper reviews the current status of the integration of RS technology with multi-source data, machine learning (ML), and PBMs for water quality monitoring, modeling, and management, along with practical applications. It offers a thorough analysis of their advantages and challenges, identifies the current research gaps, and outlines future research directions. The goal is to enhance the role of integrated methods in improving water quality in aquatic ecosystems, support sustainable water resource management, and strengthen scientific decision-making in the face of climate change and growing anthropogenic pressures.

1. Introduction

Climate change has altered precipitation patterns, intensified extreme weather events, and raised global temperatures. These factors collectively exert significant impacts on both water quantity and quality [1]. In parallel, land use and land cover changes (LUCC), driven by urbanization, agricultural expansion, industrial development, and forest fires, have further deteriorated water quality. These combined stressors not only compromise the integrity of global aquatic ecosystems, but also exacerbate water scarcity and pollution, posing serious threats to ecosystem stability and the safety of drinking water supplies [2,3,4]. In this context, robust water quality monitoring, predictive modeling, and adaptive management have become essential for mitigating the compounded impacts of climate change and human activities, thereby supporting the sustainable use and protection of global water resources.

We conducted a review of water quality monitoring and modeling approaches published over the past ten years. The 2015–2024 timeframe was chosen as it marks both the rapid growth of publications and the maturation of AI techniques in remote sensing, reinforced by the policy impetus of SDG 6, which emphasizes water quality monitoring. Relevant studies were identified through a keyword-based search strategy employing terms such as “hydrology”, “aquatic environment monitoring”, “water quality”, “water quality monitoring”, “remote sensing”, “machine learning”, “process-based models”, and “integrated methods” to identify relevant studies. To ensure comprehensive coverage, broader terms such as “artificial intelligence” were also included. Web of Science was the only database used for keyword-based statistical analysis, while cited references were drawn from multiple databases. Figure 1 and

Table 1. Literature search and paper statistics.

Keywords Category	Search Strategy	Time Interval	Number of Papers Reviewed
Remote sensing Water quality	Remote sensing OR Remote sensing data AND Water quality monitoring	2015–2024	7346
Artificial intelligence Water quality	Artificial intelligence AND Water quality	2015–2024	568
Machine learning Water quality	machine learning AND Water quality	2015–2024	580
Machine learning Remote sensing Water quality	Machine learning AND Remote sensing AND Water quality	2015–2024	119
Hydrology Water quality	Hydrology OR Hydrological processes AND Water quality OR Water pollution	2015–2024	40,897
Aquatic environment monitoring	Aquatic environment AND Remote sensing OR Machine learning	2015–2024	35,315
Process-based models Water quality	Water quality AND Process-based models	2015–2024	83
Process-based models Remote sensing Water quality	Process-based models AND Remote sensing AND Water quality	2015–2024	9

provide a visual summary of this search strategy and the trend in publication numbers, thereby clarifying the scope and distribution of the reviewed literature. Details of the search strategy and the number of articles retrieved are presented in Table 1. This review highlights advancements in individual technologies such as remote sensing (RS), machine learning (ML), and process-based models (PBMs), as well as their integrated applications in water quality monitoring and modeling, establishing a foundational understanding for future research.

RS offers broad coverage. Water-quality monitoring is constrained by four complementary resolutions—spatial, temporal, spectral, and radiometric—each imposing distinct limitation. Spatial resolution determines the minimum detectable feature size, influencing the monitoring of small water bodies, while temporal resolution dictates the ability to capture short-term dynamics such as algal blooms or sediment pulses [5,6,7]. Spectral resolution is essential for retrieving optically active constituents (e.g., Chl-a, TSS, CDOM) through diagnostic bands, whereas radiometric resolution defines the sensitivity to subtle reflectance differences, particularly in dark or turbid waters [6,7,8,9,10]. In practice, trade-offs among these resolutions are often unavoidable; however, recent studies emphasize the growing importance of sensor/data fusion and advanced algorithms in mitigating these limitations [7,9]. Between 2015 and 2024, the number of review articles related to water quality monitoring and modeling has shown a steady upward trend. Specifically, 338 articles were published in 2015, followed by 428 in 2016, 418 in 2017, 513 in 2018, and 647 in 2019. The volume increased substantially in 2020, reaching 802, and continued to rise sharply in subsequent years, with 987 in 2021, 984 in 2022, 1033 in 2023, and 1196 in 2024. This sustained growth underscores the rapid expansion of research in this field and highlights the increasing reliance on remote sensing (Figure 1). In recent years, notable progress has been achieved in the application of RS technologies for water quality monitoring and prediction, particularly in the retrieval (inversion) of key parameters such as chlorophyll-a (Chl-a) concentration, suspended matter concentration, and dissolved oxygen levels [5,6]. RS has been increasingly used in agricultural regions for soil environment assessment and agricultural hydrological process analysis, in addition to crop condition monitoring [11,12,13]. RS still encounters notable limitations, including susceptibility to atmospheric interference, reduced accuracy in optically complex waters, and difficulty in estimating non-optically active constituents such as total nitrogen (TN), total phosphorus (TP), and chemical oxygen demand (COD) [9].

ML has been widely applied in water quality monitoring, forecasting, and management, covering real-time monitoring, predictive modeling, anomaly detection, pollution source identification, and process optimization. When combined with the IoT (Internet of Things), these features enable continuous data collection and real-time management [14,15,16]. However, ML faces several limitations, such as its dependence on high-quality datasets, sensitivity to feature selection, unstable optimization of complex models (e.g., ANNs and graph neural networks, GNNs), and insufficient interpretability. Data sparsity and differences in network topology also restrict the generalization capability of models, necessitating improvements through data augmentation or transfer learning [14,15,16]. As shown in Figure 1, the number of reviewed articles on the application of ML in water quality monitoring, forecasting, and management was relatively limited in the earlier years but has exhibited a marked upward trajectory over the past decade. Only 2 articles were published both in 2015 and in 2016, increasing slightly to 3 in 2017, 11 in 2018, and 19 in 2019. The volume rose to 38 in 2020 and then expanded sharply in recent years, with 72 in 2021, 88 in 2022, 158 in 2023, and 187 in 2024. This sustained growth underscores the accelerating interest in and expanding application of ML technologies in water quality research and management.

Process-based water quality models, which simulate physical, chemical, and biological processes in aquatic ecosystems like lakes and rivers, are valuable for describing changes in water quality and assessing management scenarios, such as the impacts from climate events or human activities. Notable PBMs, such as CE-QUAL-W2, EFDC, and Delft3D-WAQ, have been widely used but continue to face challenges, including limited representation of ecological interactions, high parameter uncertainty, and significant computational demands, particularly when applied to complex ecological–social systems [17,18,19]. In contrast to the steady rise in ML-related reviews, the number of review articles on PBMs has remained relatively limited and demonstrates fluctuations over the past decade. Specifically, two articles were published in both 2015 in 2016, increasing to nine in 2017, before declining to three in 2018 and seven in 2019. The volume rose slightly to 5 in 2020 and then fluctuated in the following years, with 12 in 2021, 19 in 2022, and 12 in both 2023 and 2024. Although fewer in number compared to RS- and ML-related publications, PBM studies continue to play an essential role in elucidating the mechanistic processes governing water quality dynamics, and their consistent presence underscores the enduring importance of mechanistic modeling in water quality research. Future developments should emphasize multi-scale coupling, integration of ecological mechanisms, and interdisciplinary approaches to enhance decision-support capabilities amid global environmental change. Recent efforts combine PBMs with data-driven techniques, including RS and ML, to improve data acquisition and pattern recognition, thereby increasing modeling accuracy and response efficiency.

Researchers have explored the integration of RS and ML, combining the observational advantages of RS with the predictive capabilities of ML. This integration has demonstrated enhanced accuracy in retrieving complex water quality parameters and improving model generalizability across diverse environmental contexts [20,21]. Likewise, coupling RS with PBMs allows for the assimilation of RS-derived inputs into model boundaries and calibration processes, thereby strengthening mechanistic simulation accuracy and adaptive forecasting under dynamic environmental conditions [22,23]. Nevertheless, both integration pathways continue to face persistent challenges, such as uncertainty propagation, model incompatibility, and system-level complexity.

To address these shortcomings, our review aims to articulate and evaluate an integrated RS–PBM–ML framework. By coupling multi-source observations from RS with the mechanistic foundations of PBMs and the adaptive learning capacity of ML, this leverages complementary strengths to overcome individual limitations: RS enhances coverage for PBMs and ML, PBMs provide physical and process-based understanding for ML predictions, and ML enhances calibration, forecasting accuracy, and pattern recognition for RS- and PBM-based approaches. Together, this synergy establishes a “multi-source data–mechanistic modeling–intelligent prediction” paradigm that advances monitoring, improves accuracy and interpretability, and strengthens decision support in water governance [6,7,8,9,24,25,26,27,28]. Ultimately, this work advances water quality monitoring technologies and supports evidence-based decision-making in water governance.

2. Current Developments in RS Applications for Water Quality Monitoring and Modeling

Water quality is determined by the physical, chemical, and biological characteristics of water. These attributes have been quantified primarily through on-site sampling and laboratory analyses [22,23]. But nowadays, RS has greatly transformed traditional monitoring approaches and provides more effective means for water resource management and environmental protection [29,30,31].

Table 2. Available satellites for RS water quality monitoring and their relevant parameters.

	Satellite Sensor	Launch Date	Spatial Resolution (m)	Spectral Resolution Band	Temporal Resolution (Day)	Spectrum Ranges (nm)	Country
	NIMBUS-7 CZCS	24 October 1978	825	6 bands	6	443–12,500	US
	Landsat7 ETM+	15 April 1999	15–60	8 bands	16	450–1250	US
	SeaWiFS	1 August 1999	500–1100	8 bands	1–2	412–905	US
	AVHRR	13 October 1978	1100	6 bands	1–2	580–12,500	US
	EO-1 ALI	21 November 2000	10–30	10 bands	16	433–2350	US
Multi-spectral	WorldView-2	8 October 2009	0.46–0.52	9 bands	1.1	400–1040	US
	MERIS	March 2002	300	15 bands	1–3	407.5–905	EU
	MODIS	18 December 1999	250–1000	36 bands	1	405–14,385	US
	Landsat-8 OLI	11 February 2013	15–30	9 bands	16	435–2294	US
Hyper-spectral	Hyperion	21 November 2000	30	242 bands	16	349.896–2582.28	US
	HJ-1A HSI	6 September 2008	100	115 bands	4	450–950	CN
Sensor for UAV	GaiaSky-mini	-	0.04	176 bands	/	400–1000	CN

lists several satellites and their specifications pertinent to water quality monitoring, many of which feature broad spectral ranges.

2.1. Applications of RS in Water Quality Monitoring and Prediction

In the theory of water quality RS monitoring, the estimation of water quality parameters primarily relies on the spectral information of water body components. These water quality parameters can be classified into two categories: one comprising optically active constituents (OACs), including Chl-a, total suspended solids (TSS), CDOM, and water transparency (Secchi Disk Depth, SDD); and the other comprising non-optically active constituents (NOACs), such as TN, TP, DO, COD, and BOD [29]. OACs interact with light through absorption, refraction, and scattering, while NOACs, due to their lack of distinct optical properties, cannot be directly measured using spectral methods. Currently, the commonly employed methods for the inversion of water quality parameters include empirical, semi-empirical, analytical, and ML approaches [32]. This section primarily reviews the application of satellite RS in monitoring several typical water quality parameters. It summarizes four optically active water quality parameters commonly derived from satellite data, as presented in Table 3.

• Optically Active Constituents

The concentration of Chl-a is a pivotal indicator for assessing eutrophication and algal blooms in aquatic ecosystems. RS technology, by capturing the absorption characteristics of chlorophyll in the green and red light wavelengths, utilizes algorithms to invert Chl-a concentrations [9]. Satellite RS, with its extensive coverage, is particularly suited for tracking the macro-dynamics of algal blooms. Pahlevan et al. (2020) employed Landsat-8 and Sentinel-2 data, applying ML techniques to invert the Chl-a concentration of the Great Lakes in the United States, achieving a precision of R² = 0.78 and revealing the seasonal variations of algal blooms [30]. In contrast, drone-based RS excels in capturing the spatial heterogeneity of algal blooms at high resolution. Liu and his team utilized drone-based hyperspectral sensors to monitor the blue-green algal blooms in Taihu Lake, achieving a spatial resolution of 0.05 m and precisely identifying the spatial distribution of bloom species [33]. The combination of both methods significantly enhances monitoring efficiency. Markogianni and colleagues, by integrating Sentinel-2 satellite data with drone imagery, monitored the Chl-a distribution in Greece’s Polyphytos Reservoir. The satellite data revealed the seasonal patterns of the blooms, while the drone data provided detailed characteristics of high-concentration areas, thus optimizing early-warning strategies for blooms [34]. Additionally, Guo and his team, combining drone hyperspectral data with MODIS satellite imagery, successfully captured both the short-term dynamics and long-term trends of Chl-a concentrations in China’s Chaohu Lake [35].

The concentration of TSS is a crucial indicator for assessing water turbidity and ecological health. RS technology, by leveraging the scattering and absorption characteristics of suspended particles, can establish inversion models using multispectral or hyperspectral RS imagery to estimate TSS concentrations [36]. In satellite RS, Liu et al. (2017), utilizing Sentinel-2 data, successfully inverted the TSS concentration of Poyang Lake in China with an empirical regression model, achieving a model explanation rate of up to R² = 0.93, thereby validating the high applicability of Sentinel-2 in turbid aquatic environments [37]. Additionally, Mo et al. (2024) [38] integrated Sentinel-2 multispectral imagery with various ML algorithms to invert the suspended particulate matter (SPM) concentration at the estuary of the Pinglu Canal in China, discovering that the XGBoost model performed best (R² = 0.90). This study revealed the spatial variation of SPM, which was influenced by both tidal and anthropogenic disturbances [38]. This research offers an effective approach for using moderate-resolution RS data to monitor TSS in estuarine and coastal waters.

The transparency (SDD) and colorimetric properties (CDOM) of water are pivotal indicators of the optical characteristics and pollution status of aquatic environments. RS technology, by analyzing the reflection properties in the blue and green light wavelengths, can effectively estimate the SDD and CDOM values [39]. In regional- or global-scale monitoring, Olmanson et al. (2008), utilizing MODIS data, inverted the SDD of global lakes and constructed a 40-year timeseries, revealing the long-term effects of climate change on lake transparency [40]. In nearshore or small-scale complex aquatic environments, drone-based RS offers superior spatial resolution. Keller and his team used hyperspectral drone data to monitor the CDOM concentration in the waters along the Baltic Sea coast, achieving an accuracy of R² = 0.82 [41]. Liu and others further integrated Sentinel-2 and drone data to systematically monitor the spatial distribution of SDD and CDOM in Dongting Lake, China, identifying regions influenced by pollution sources and providing critical support for water quality management [42]. Similarly, Lin et al. (2021), through the fusion of multispectral drone imagery and Landsat-8 data, studied the long-term variation of CDOM concentrations in Dianchi Lake, revealing the impact of agricultural runoff [43]. Furthermore, Kutser and his team, utilizing MERIS data, estimated the concentrations of CDOM and dissolved organic carbon components in Swedish lakes, confirming the feasibility and broad applicability of RS technology in monitoring the spatiotemporal variations of CDOM [44].

• Non-Optically Active Constituents

NOACs, such as TN, TP, and DO, do not directly influence the spectral reflection properties of water bodies, rendering traditional RS methods inadequate for their direct concentration estimation. However, these parameters often exhibit strong ecological or statistical correlations with optically active factors (such as Chl-a and SPM), providing a theoretical foundation for their indirect inversion through RS. In recent years, with the integration of multispectral RS and ML technologies, significant advancements have been made in the RS inversion of NOACs. Studies have demonstrated that by constructing ML models and integrating Sentinel-2 data, parameters such as COD_mn, DO, TN, and TP can be estimated with moderate precision, with the best model achieving an R² of 0.52 [45]. To enhance model performance in complex aquatic environments, Zhu et al. (2023) proposed a zonal modeling strategy that significantly improved the fitting accuracy of TN and TP in regions of local heterogeneity, with all parameters achieving R² values exceeding 0.5 [46]. Addressing the spatial heterogeneity issue in large-scale lake environments, Zhang et al. (2022) introduced a geographic partition-based RS modeling strategy to improve the monitoring accuracy of TN and TP in expansive lake settings [47]. Despite the physical limitations of NOACs in RS inversion, the application of statistical modeling, regional partitioning, and ML techniques has still enabled the attainment of highly accurate estimations in spatially extensive and hydrologically diverse contexts, thereby providing crucial support for water quality monitoring and watershed management. The retrieval model algorithms for Chl-a, TSS, SDD, and CDOM are shown in Table 3.

Table 3. Retrieval model algorithms for Chl-a, TSS, SDD, and CDOM.

Satellite (Sensor)	Water Quality Parameter	Retrieval Algorithm Formula	References
Landsat-8 OLI	TSS	$\log \left(TSS\right)=1.5212\left({\displaystyle \frac{\log Rrs\left(480\right)}{\log Rrs\left(560\right)}}\right)-0.3698$	[24]
Landsat-8 OLI, GF-1 WFV	Chl-a	$\log \left(Chl-a\right)=1.613\left({\displaystyle \frac{\log Rrs\left(480\right)}{\log Rrs\left(655\right)}}\right)-1.0718$$\mathrm{l}\mathrm{n}\left(Chl-a\right)=487/B4-0.86$	[24,25]
Advanced Land Imager (ALI)	CDOM	$a_{CDOM}\left(420\right)=5.20·{\left(B2/B3\right)}^{-2.76}$	[44]
Landsat-8 OLI, Landsat-5 TM	SDD	$\ln \left(SDD\right)=b_0+b_1\left(B2/B4\right)+b_2·B1$$\ln \left(SDD\right)=-4.064+1.961\left(B1/B3\right)+166.4\left(B7\right)$	[26,27]

$a_{CDOM}$(420): CDOM Absorption Coefficient, Rrs (λ): represents the reflectance at a specific wavelength λ, used for inverting water body parameters.

Table 3. Retrieval model algorithms for Chl-a, TSS, SDD, and CDOM.

Satellite (Sensor)	Water Quality Parameter	Retrieval Algorithm Formula	References
Landsat-8 OLI	TSS	$\log \left(TSS\right)=1.5212\left({\displaystyle \frac{\log Rrs\left(480\right)}{\log Rrs\left(560\right)}}\right)-0.3698$	[24]
Landsat-8 OLI, GF-1 WFV	Chl-a	$\log \left(Chl-a\right)=1.613\left({\displaystyle \frac{\log Rrs\left(480\right)}{\log Rrs\left(655\right)}}\right)-1.0718$$\mathrm{l}\mathrm{n}\left(Chl-a\right)=487/B4-0.86$	[24,25]
Advanced Land Imager (ALI)	CDOM	$a_{CDOM}\left(420\right)=5.20·{\left(B2/B3\right)}^{-2.76}$	[44]
Landsat-8 OLI, Landsat-5 TM	SDD	$\ln \left(SDD\right)=b_0+b_1\left(B2/B4\right)+b_2·B1$$\ln \left(SDD\right)=-4.064+1.961\left(B1/B3\right)+166.4\left(B7\right)$	[26,27]

$a_{CDOM}$(420): CDOM Absorption Coefficient, Rrs (λ): represents the reflectance at a specific wavelength λ, used for inverting water body parameters.

• Delineation of Water Boundaries and Monitoring of Changes

RS image processing techniques commonly employ methods such as threshold segmentation, supervised classification, and unsupervised classification to distinguish water bodies from non-water regions. Among these, the Normalized Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI), the Normalized Difference Moisture Index (NDMI), the Automated Water Extraction Index (AWEI), and the Water Index (WI) are widely applied due to their effectiveness for water body identification and monitoring. These indices have been widely applied across diverse environments. Specifically, NDWI and MNDWI are particularly effective in distinguishing water bodies from vegetation and built-up areas; NDMI is valuable for detecting water content and wetlands; and AWEI and WI demonstrate strong robustness in automatically extracting water under shadow or urban conditions. The corresponding parameters are summarized in Table 4.

Table 4. Common parameters for delineating water bodies.

Index	Full Name	Formula	Source
NDWI	Normalized Difference Water Index	NDWI = (Green − NIR)/(Green + NIR)	Index DataBase
MNDWI	Modified Normalized Difference Water Index	MNDWI = (Green − SWIR)/(Green + SWIR)	Index DataBase
NDMI	Normalized Difference Moisture Index	NDMI = (NIR − SWIR)/(NIR + SWIR)	Index DataBase
AWEI	Automated Water Extraction Index	AWEI = 4(Green − SWIR1) − (0.25 × NIR + 2.75 × SWIR2); AWEI = Blue + 2.5 × Green − 1.5(NIR + SWIR1) − 0.25 × SWIR2	Feyisa et al. (2014) [48]
WI	Water Index	WI = (Blue + Green)/(NIR + SWIR)	Fisher et al. (2016) [49]

Table 4. Common parameters for delineating water bodies.

Index	Full Name	Formula	Source
NDWI	Normalized Difference Water Index	NDWI = (Green − NIR)/(Green + NIR)	Index DataBase
MNDWI	Modified Normalized Difference Water Index	MNDWI = (Green − SWIR)/(Green + SWIR)	Index DataBase
NDMI	Normalized Difference Moisture Index	NDMI = (NIR − SWIR)/(NIR + SWIR)	Index DataBase
AWEI	Automated Water Extraction Index	AWEI = 4(Green − SWIR1) − (0.25 × NIR + 2.75 × SWIR2); AWEI = Blue + 2.5 × Green − 1.5(NIR + SWIR1) − 0.25 × SWIR2	Feyisa et al. (2014) [48]
WI	Water Index	WI = (Blue + Green)/(NIR + SWIR)	Fisher et al. (2016) [49]

Pekel et al. (2016), utilizing Landsat data and the NDWI index, constructed a global water body distribution atlas, achieving an impressive identification accuracy of 95%, thereby establishing a robust data foundation for global water resource assessments and disaster monitoring [28]. In complex urban aquatic environments, Wang and his team, using multispectral data acquired by drones combined with the Extreme Gradient Boosting (XGBoost) algorithm, successfully identified black and odorous water bodies in typical urban districts of Guangzhou, significantly enhancing both monitoring precision and efficiency and offering a highly cost-effective solution for the rapid identification of urban water pollution [50]. Concerning ecological changes at the watershed scale, Wu and Liu, based on long-term RS timeseries data, systematically analyzed the impact of the Three Gorges Project on water levels and wetland patterns in Poyang Lake, revealing significant changes in lake connectivity and ecological configuration under the influence of hydraulic engineering [51]. Giardino and his team further introduced airborne hyperspectral RS to quantitatively monitor suspended particulates and aquatic vegetation distribution in Italy’s Lake Trasimeno, shedding light on how vegetation changes regulate water transparency based on water body classification [52].

The advantages of multi-temporal RS in monitoring the dynamic changes of water bodies have become increasingly evident, enabling the effective observation of changes in water area, shape, and connectivity, and revealing the cumulative impacts of climate change and human activities. In the study of wetland connectivity, Ashok and his team, using Landsat imagery, developed a combined monitoring system using NDVI and NDWI to quantify the impact of urban expansion and hydrological fluctuations on the wetland structure of the western Himalayas in India [53]. For localized, high-frequency monitoring needs, Xie et al. (2024) [54] employed a consumer-grade drone RGB camera to invert water turbidity in small lakes within the Taihu Basin. Their results demonstrated that this method, even under low-cost conditions, still exhibited strong spatial resolution and accuracy [54]. Moreover, Lobo and his team, using a 40-year Landsat timeseries, assessed the changes in suspended particulate matter in the Tapajos River of the Amazon caused by gold mining activities, validating the value of RS in monitoring long-term water quality evolution [55].

• Analysis of Water Quality Trends and Monitoring of Emergencies

RS technology provides formidable support for long-term water quality monitoring. By leveraging decades of satellite data, researchers are able to precisely analyze the changing trends of water quality parameters, including key indicators such as Chl-a concentration and suspended matter levels. These variations in water quality directly influence the health assessment of water bodies and the scientific management of water resources. Tanner et al. (2022) utilized Landsat data to examine the fluctuations in Chl-a concentrations in Utah Lake, uncovering the impacts of climate change, urban expansion, and agricultural runoff on water quality [56]. This study employed the non-parametric Mann–Kendall test to assess the trends of Chl-a in different regions of Utah Lake, revealing that water quality in most areas has declined, particularly in regions heavily affected by development and population growth. However, the accuracy of Chl-a inversion may be influenced by seasonal variations in algal biomass, which warrants further investigation. Furthermore, research in the Hawke’s Bay region further validates the effectiveness of RS technology in long-term water quality monitoring. By analyzing Chl-a concentrations from 2018 to 2023 using the Cawthron Institute’s RS algorithms, the study found that the region’s water quality fluctuated in response to seasonal changes and climatic cycles [57]. This result highlights the capacity of RS to offer long-term data support, aiding in the identification of potential causes behind water quality changes.

Moreover, its application in monitoring sudden water quality pollution events is equally indispensable. By combining hyperspectral and thermal infrared RS, researchers can efficiently track the extent and severity of pollution, particularly in events such as oil spills. Yang et al. (2023) proposed a method that integrates hyperspectral and thermal infrared RS technologies, analyzing the spectral characteristics of various oil types to accurately identify oil slicks on the water surface and assess the influence of slick thickness on the spread of pollution [58]. Studies indicate that this technological combination significantly enhances the accuracy of oil pollution detection, especially in complex marine environments where RS provides real-time monitoring data to support emergency response efforts. Duan et al. (2022) developed an unsupervised learning-based oil pollution detection method, merging the Isolation Forest algorithm with hyperspectral data [59]. This approach successfully improved detection accuracy, particularly in the presence of noise interference, by automatically eliminating noise, thereby enhancing the reliability of the results [59].

2.2. The Trend of Multi-Technology Integration in Assessing Water Quality

RS continues to confront multiple technological challenges in practical applications. These primary obstacles include the insufficient adaptability of atmospheric correction algorithms, the difficulty of balancing spatial and temporal resolution, limited model accuracy in optically complex waters, and the weak RS estimation capabilities for non-optical parameters such as TP, TN, and COD [10,60]. Moreover, environmental factors such as cloud and fog interference, signal attenuation, and surface water fluctuations further compromise data availability and the stability of results. The RS inversion process inherently possesses considerable uncertainty, particularly in highly turbid or composite-polluted waters, where fluctuations in key parameters such as Chl-a and TSS directly affect the reliability and stability of model simulations [60].

With the continuous advancement of RS in water quality monitoring, its integration with ML and PBMs has emerged as a crucial direction for enhancing the accuracy of water quality inversion and the system’s simulation capabilities. This technological fusion not only compensates for the shortcomings of single-method approaches when handling high-dimensional and complex environmental data, but it also expands the applicability of RS in addressing dynamic, multi-factorial-driven changes in water quality.

• The Coupling of RS and ML

The integration of RS and ML offers more robust and generalized solutions for the automatic inversion of water quality parameters, temporal predictions, and the response to pollution events. Numerous studies have demonstrated that algorithms such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and random forests (RFs) can effectively capture complex nonlinear relationships in RS imagery, thereby achieving high-precision estimations of key indicators such as Chl-a, turbidity, and TSS [21]. Mohan et al. (2025), after systematically reviewing 186 studies, noted that ML can not only manage the high dimensionality and redundancy of spectral data but also facilitate model transfer and generalization across different RS products [20]. For further details, see Section 3 on the integration of RS and ML.

• The Coupling of RS and PBMs

RS technology provides high-resolution initial boundary conditions, driving data, and dynamic calibration references for hydrodynamic, water quality, and ecological models, thus advancing the shift from “monitoring–inversion” to “prediction–management” applications [7]. Ye et al. (2019) noted in their review that optically active constituents (e.g., Chl-a, CDOM) derived from RS can serve as input variables for one-dimensional or two-dimensional water quality models, while NOACs (e.g., TP, TN, DO) can be indirectly simulated via remotely sensed, model-driven intelligence [61]. This approach mitigates the reliance of traditional physical models on high-quality spatiotemporal data, enhancing modeling efficiency and regional adaptability. Simultaneously, PBMs provide mechanistic explanations and trend validation for RS monitoring results. For instance, by integrating remote observations with reaction–transport models, the spatiotemporal evolution of pollutants and nutrient response processes can be traced, offering scientific support for eutrophication management [62]. Some studies have already begun exploring data assimilation mechanisms in coupled models (e.g., Soil and Water Assessment Tool, SWAT, hydrological–ecological models) using RS data, achieving an initial information loop between RS, models, and management. For further details, see Section 4 on the coupling of RS and PBMs.

The integration of RS with ML and PBMs is constructing a new generation of water quality monitoring systems centered around “multi-source, heterogeneous data-driven, mechanistic modeling, and intelligent prediction.” The advantage of this system lies in its combination of real-time capabilities, accuracy, and scalability, positioning it as a critical technological path for addressing global climate change and regional water environment challenges in the future.

3. Leveraging ML in RS for Water Quality Monitoring: From Data Processing to Predictive Accuracy

ML algorithms are ideally suited to address the complex nonlinear relationships between RS data and water quality parameters. By integrating RS with ML techniques, predictive models can be developed to enhance both the accuracy and efficiency of water quality assessments. These models can identify and predict trends in water quality, facilitating timely interventions to safeguard aquatic ecosystems [63]. As a branch of artificial intelligence, ML enhances decision-making by allowing computers to learn from data and patterns. In the context of water quality monitoring and simulation, ML applications primarily focus on three key areas: (1) the processing and analysis of water quality monitoring data; (2) the modeling and prediction of water quality parameters; and (3) the optimization and integration of models, including the fusion of diverse technologies to support more comprehensive and intelligent water quality assessments.

3.1. Processing RS Data Using ML Techniques

ML algorithms such as clustering analysis and interpolation methods are widely applied in water quality data preprocessing. Clustering algorithms have proven effective in identifying and removing outliers, while interpolation techniques are commonly employed to fill missing data gaps. For instance, in their study on water quality monitoring, Li et al. applied clustering algorithms to preprocess RS images, successfully eliminating outliers in water reflectance data. This significantly improved data quality and provided a reliable foundation for the accurate inversion of water quality parameters [64].

RS monitoring data often comprises a large number of parameters and variables, many of which may have limited relevance for water quality evaluation and prediction. To address this, feature selection algorithms within machine learning, such as Recursive Feature Elimination (RFE) and PCA, play a crucial role in this process. These techniques help identify and extract the most influential features, reducing data dimensionality while improving both the efficiency and predictive accuracy of water quality models. Zou et al. applied PCA to extract key features from RS imagery, effectively reducing data volume while enhancing both the training efficiency and predictive performance of water quality parameter inversion models. This approach significantly improved the model’s ability to detect and respond to changes in water quality [65]. Moreover, feature selection algorithms also facilitate a deeper understanding of the primary factors influencing water quality changes, such as Chl-a concentration and suspended solids levels.

3.2. Modeling and Prediction of Water Quality Parameters Driven by ML Techniques

In recent years, researchers have increasingly introduced ML and deep learning (DL) models to enhance the analytical efficiency and predictive accuracy of RS data. These advancements have significantly expanded the applicability of RS, particularly in multi-source data fusion and cross-temporal scale modeling [20,66].

In the prediction of DOC, Mandal et al. constructed a random forest and residual neural network model integrating Landsat RS bands and spatiotemporal covariates, achieving a coefficient of determination (R²) of 0.83 and a 50% reduction in Root Mean Square Error (RMSE), significantly outperforming traditional spectral-based algorithms. This demonstrates the superior capability of nonlinear models in capturing the spatial heterogeneity of DOC [67]. In monitoring urban surface water pollution, Do et al. utilized multi-source Sentinel-1/2 RS imagery combined with the Cubist model to estimate concentrations of TSS, COD, and BOD in Hanoi, Vietnam. This approach effectively replaced expensive ground-based sampling while accommodating the complexity of urban landscapes and the dynamic variability of pollutant distributions [68].

The fusion of multi-source information is also a vital direction in the integration of RS and ML. Kupssinskü et al. integrated Sentinel-2 and drone imagery, applying algorithms such as neural networks and random forests to predict Chl-a and suspended solids concentrations. The model achieved an accuracy of R² > 0.8 across different spatial resolutions, validating its robust adaptability and cross-scale stability [69]. Li et al. systematically compared various ML algorithms in water boundary extraction, finding that models such as SVM, XGBoost, and neural networks demonstrated greater resistance to interference compared to the MNDWI index method, particularly in complex backgrounds like building shadows and mountainous terrains, maintaining strong transfer and generalization capabilities [70].

Furthermore, deep learning models (e.g., CNN, LSTM, and their hybrid architectures) are highly effective in capturing temporal changes, predicting water quality trends, and providing early warnings for emergent events like algal blooms. These models significantly enhance the representation of nonlinear relationships and improve the ability to model complex spatiotemporal dependencies in water quality data [21]. Deng et al. (2024) also highlighted that deep learning-based RS intelligent recognition methods such as U-Net are shaping a new paradigm of “ubiquitous sensing–intelligent analysis” [21]. This emerging approach holds great potential for supporting the construction of next-generation global water resource monitoring systems, particularly in challenging environments characterized by multiple pollution sources, high optical turbidity, and low-light observation conditions [66].

In the domain of non-optical RS, Leggesse et al. integrated Synthetic Aperture Radar (SAR) data with ML for water quality modeling in highly turbid lakes, effectively overcoming the limitations of visible light data [71]. For example, during the 2021 Ulva prolifera bloom in the Southern Yellow Sea, Sentinel-1 SAR imagery was integrated with a random forest model to enable automatic detection and mapping of floating algal mats. The all-weather, day-and-night capability of SAR effectively addressed the limitations of cloud cover and precipitation in monitoring algae [72], while the random forest classifier maintained stable detection accuracy in complex nearshore environments. Collectively, these findings demonstrate the strong potential of integrating SAR with ML approaches for continuous algal bloom and water quality monitoring, even under extreme weather conditions or in highly turbid waters [73]. Similarly, Gnann et al. successfully applied near-surface RS combined with CNN to detect microplastics on water surfaces, thereby extending the applicability of RS in pollutant identification [74]. Additionally, Li et al. (2024) [75], addressed the challenge of water depth inversion in turbid waters, such as harbors, by innovatively combining geographically weighted regression (GWR) with random forest models. This approach significantly improved prediction accuracy and stability, demonstrating the enhanced potential of integrating spatial information with machine learning techniques [75].

4. Integrating RS and PBMs for Enhanced Understanding of Water Quality Dynamics and Mechanisms

As water environmental challenges grow increasingly complex, PBMs are playing an increasingly critical role in bridging monitoring data with a mechanistic understanding of water quality dynamics. Serving as foundational tools in modern water quality monitoring and early warning systems, PBMs use mathematical formulations and dynamic process rules to simulate the evolution of physical, chemical, and ecological processes within aquatic environments. This capability supports a wide range of applications, including pollutant identification, source tracking, and the prediction of short-, medium-, and long-term water quality trends, ultimately informing more effective management and decision-making.

The profound integration of RS technology with process-based water quality simulation models is progressively shaping a monitoring system characterized by multi-scale coverage, high interpretability, and strong generalizability. Integrating PBMs with RS combines the spatial richness of observational data with the explanatory power of mechanistic modeling, providing robust support for high-resolution simulations and the dynamic management of complex water environments [76].

4.1. Application of PBMs in Water Quality Simulation

PBMs facilitate the dynamic simulation of internal responses in water bodies by characterizing the mechanisms behind changes in key water quality parameters, such as DO and nutrient concentrations. For example, the variation in DO is influenced by multiple factors, including temperature, water flow, biological metabolism, and atmospheric reoxygenation. Li et al. (2013) [76], in their empirical study in the Wenruitang River Basin, constructed a BOD-DO model based on continuous longitudinal monitoring data, successfully inverting process parameters such as oxygen consumption coefficients, source loads, and self-purification capacity in urban rivers. This provided an accurate mechanistic foundation for dynamic DO simulation [76]. Such studies offer vital references for the management of other eutrophic and hypoxic water bodies. Additionally, Lemaire et al. (2021) developed a system dynamics model that coupled DO with nutrients, temperature, and other indicators, revealing the mechanism of DO decline caused by intensified heterotrophic respiration under high summer temperatures and emphasizing the importance of the collaborative expression of temporal and spatial heterogeneity in scientific simulations [77].

Some well-known PBMs, such as Delft3D, have been widely applied in water quality simulation and prediction. For instance, Liu et al. utilized the Delft3D model to evaluate the impact of various water resource allocation schemes on water quality in Chagan Lake, providing a scientific basis for achieving the management targets of total nitrogen and total phosphorus [78]. Li et al. employed the Delft3D model to simulate the spatiotemporal dynamics of cyanobacterial blooms in a tropical reservoir and validated the model’s accuracy by comparing it with field observation data, thereby offering reliable predictive outcomes for water quality research and management [79]. Additionally, Tong et al. (2022) [80] integrated the Delft3D model with data-driven models such as artificial neural networks (ANNs) and random forests (RFs) to predict the spatiotemporal distribution of emerging contaminants (ECs) in a tropical reservoir. By using the water quality data generated by the Delft3D model as input for training the data-driven models, the accuracy and efficiency of the predictions were enhanced [80].

In the context of pollutant transport and transformation within water bodies, PBMs integrate input sources, flow pathways, hydrodynamic characteristics, and reaction mechanisms to reproduce the spatial distribution patterns and concentration evolution trends of pollutants. Tong et al. (2022) [81] systematically reviewed the current status of emerging contaminant (EC) modeling, noting that widely used PBMs such as WASP and SWAT can now simulate the transport, adsorption, and degradation processes of pollutants like Pharmaceutical and Personal Care Products (PPCPs), Per-and polyfluoroalkyl substances (PFASs), and Endocrine Disrupting Chemicals (EDCs) in river and lake systems. They emphasized the importance of models expressing multi-phase reactions and interactions [81].

In practical applications, Stonewall et al. (2019) [82] employed the Stochastic Empirical Loading and Dilution Model (SELDM) to simulate water quality changes in river systems after urban runoff pathways. Their findings indicated that, even in the absence of high-density monitoring, the model effectively predicted pollutant loads and risk zone distribution in multiple scenarios, demonstrating its strong capacity for policy support [82].

When integrated with climate change factors, PBMs acquire the ability to simulate short-, medium-, and long-term water quality trends. Wang et al. (2022) [83] developed an LSTM-based water temperature prediction system that successfully achieved high-precision long-sequence simulations in deep-water reservoirs. This model not only significantly enhanced computational efficiency but also seamlessly integrated with scheduling decision models, providing vital technical support to address ecosystem fluctuations amid global warming [83]. Similarly, Jiang et al. (2021) proposed an extended Bayesian network model that incorporated climate variables, biological activity, and management measures to dynamically predict cyanobacterial bloom risks, integrating mechanisms for handling nonlinearity and uncertainty, thus offering quantitative support for bloom control and water source safety management [84].

PBMs exhibit multi-dimensional application value in water quality prediction, revealing the inherent evolutionary mechanisms of aquatic systems while providing theoretical and decision-making tools for pollution control and ecological risk management. As modeling methods advance towards multi-source coupling and intelligent prediction, PBMs are set to play an even more central role in future water resource protection and management.

4.2. Integrated Methods for Water Quality Monitoring and Retrieval

In practical applications, RS products such as MODIS and Sentinel-2, are commonly used to retrieve water quality parameters including temperature, TSS, and Chl-a. These data are often incorporated as input variables or boundary conditions in PBMs, enabling spatiotemporal forcing and dynamic calibration of the simulation system for improved accuracy and representativeness. For example, Chen et al. integrated MODIS-derived TSS data with the Estuarine, Coastal Ocean Model with Sediment Transport (ECOMSED) model, improving the model’s three-dimensional simulation accuracy and timeliness in representing particulate matter distribution in the coastal waters of the Bohai Sea [85]. Similarly, Mano et al. successfully reconstructed the three-dimensional temperature and water quality structure of Finnish lakes by assimilating Sentinel-2 data into the Coupled Hydrodynamical Ecological model for Regional Shelf seas (COHERENS) [86]. In the North Sea and Baltic Sea regions, Pleskachevsky et al. and Kõuts et al. developed RS-based models for particulate transport and algae dispersion, achieving high-precision predictions of coastal pollution dynamics [87,88].

The fusion of RS with PBMs still faces several challenges in large-scale, systematic deployment. The inconsistency in spatial and temporal resolution between RS data and PBMs often becomes a bottleneck in model-driven accuracy. Studies show that while MODIS offers daily coverage, its spatial resolution of 250–500m is insufficient for fine-scale modeling needs; on the other hand, high-resolution RS products such as Sentinel-2 and Landsat-8 are constrained by revisit cycles and meteorological conditions, limiting their ability to ensure dynamic continuity [89]. Moreover, the RS inversion process itself introduces significant uncertainty, as fluctuations in the precision of key parameters directly impact the reliability and stability of model simulations [60]. PBMs often involve complex initial conditions and boundary control parameter settings, necessitating high-quality ground monitoring data, which increases the cost of model development and transfer applications [7].

By incorporating RS-derived data to support the definition of model boundary conditions and internal process controls, it becomes possible to effectively compensate for the scarcity of ground-based data and the challenges of long-term monitoring. More importantly, this integrated approach preserves the physical consistency of process-based models while enhancing their adaptability to external disturbances—such as floods or sudden pollution events. As a result, it demonstrates strong practicality and generalizability in complex aquatic environments, including estuaries, lakes, and bays [85,88].

5. A Synergistic Framework Integrating ML, RS, and PBMs for Enhanced Water Quality Management

At present, the integration of RS, ML, and PBMs is gradually becoming a significant trend across disciplines such as earth system science, ecological and environmental monitoring, and agricultural resource management. RS technology provides large-scale, multi-temporal spatial observation capabilities; ML methods enable advanced nonlinear modeling and feature extraction; and PBMs offer a mechanistic understanding of the physical, chemical, and biological processes within environmental systems. The synergistic combination of these three approaches not only improves the accuracy of parameter inversion and variable prediction but also enhances model interpretability and introduces knowledge-based constraints, thereby deepening insights into complex natural processes.

5.1. Insights into the Mechanisms Underlying Integrated Modeling Approaches

The integration of RS, PBMs, and ML presents a novel and powerful approach for modeling complex aquatic environments. The core of this synergistic fusion lies in leveraging RS data for large-scale observational input or auxiliary validation, employing ML as a nonlinear mapping and residual correction tool, and utilizing PBMs to impose physical constraints and mechanistic interpretability. Together, these elements form a closed-loop modeling framework that connects data, patterns, and process mechanisms. Specifically, RS provides spatially extensive and temporally continuous observations; PBMs incorporate physical and biogeochemical processes to guide model structure and ensure consistency; and ML enhances predictive performance by capturing nonlinear relationships and optimizing residuals. Figure 2 conceptually illustrates this integration process and the underlying mechanisms, reinforcing the central theme of the paper and highlighting the key research focus. The resulting hybrid models are not only physically consistent and highly predictive, but also scalable and adaptable, enabling the collaborative advancement of data-driven and mechanistic modeling paradigms [90,91].

Some studies have developed hybrid Hydro-Biogeochemical (HBGC–CNN) models that incorporate process-based model outputs (e.g., nutrients and temperature) along with RS-derived Chl-a results as input variables. This integrated approach has achieved high-precision predictions and accurate spatiotemporal reconstructions of Chl-a concentrations in coastal regions [92]. In parallel, lake water temperature modeling has begun integrating deep learning techniques with lake energy balance models, significantly enhancing the models’ responsiveness to dynamic thermal fluctuations and improving the ability to capture complex heat exchange processes [93].

5.2. Innovative Applications of Integrated Methods in Water Quality Monitoring and Cross-Domain Research

In ecosystem carbon flux simulations, researchers have increasingly coupled RS data with PBMs (such as carbon cycle models), while incorporating Bayesian machine learning algorithms to optimize model parameters and quantify associated uncertainties. This integrated approach significantly enhances the accuracy and scalability of carbon dynamics estimation in grassland ecosystems [94]. In the agricultural sector, integrated systems combining RS imagery, crop growth PBMs (such as DSSAT, Decision Support System for Agrotechnology Transfer), and ML algorithms (such as XGBoost) have been employed to predict crop yield and evapotranspiration under drought conditions. This significantly enhances the scientific nature of regional agricultural water resource management [95]. Moreover, in the retrieval of land surface temperature, scholars propose a mechanism-learning coupling framework that integrates physical radiative transfer models with deep neural networks. This approach balances the accuracy and physical consistency of RS retrieval, providing a key technological support for global climate research [96].

In the field of RS image processing, several studies combine model-driven and data-driven deep networks to carry out image restoration and fusion tasks. By co-designing physical priors and neural structures, these methods enhance the detail retention and noise robustness in RS images [97]. Going a step further, some scholars have proposed a “mechanism-learning coupling paradigm,” emphasizing that the integration of remote sensing, machine learning, and process-based modeling represents not only a methodological integration but a fundamental shift in cognitive approach. This paradigm embeds mechanistic constraints within data-driven models while incorporating observational data into mechanistic simulations, ultimately forming a highly credible and unified framework for perception–cognition–prediction. Such an approach aims to enhance interpretability, reliability, and the generalization capacity of environmental modeling systems [92].

However, research on integrated models in water quality management remains relatively limited, highlighting a critical area for future exploration. In watershed-scale water quality prediction, some studies have utilized RS imagery to extract variables such as water temperature, surface reflectance, and water body indices and combined these features with PBMs to simulate watershed hydrological processes. ML models are subsequently employed to correct residuals or directly predict key water quality parameters, including Chl-a and total suspended solids. This integrated approach offers promising potential for improving the accuracy and scalability of water quality forecasting across spatial and temporal scales. Wang et al. (2016) developed an integrated framework combining process-based water quality models with data assimilation approaches that significantly improved the forecasting of key indicators such as salinity and dissolved oxygen in Singaporean coastal waters [98]. Building upon this, Wang et al. (2019) proposed a comprehensive integrated catchment-scale monitoring and modeling approach, incorporating hydrodynamic and water quality models with data assimilation, thereby enabling both long-term predictions and real-time forecasting of water quality [99]. More recently, Tong et al. (2022) demonstrated that coupling general water quality indicators with machine learning within a hybrid process–data-driven framework can effectively predict emerging contaminants in a tropical reservoir, underscoring the potential of such synergistic approaches for advancing water quality management [80]. Together, these studies illustrate the promise of hybrid modeling strategies in overcoming the limitations of standalone approaches, thereby positioning them as powerful tools for future water quality forecasting and management. The “data–model–algorithm” three-dimensional fusion approach has been successfully applied in regions such as the lakes of the American Midwest and the Yellow River Delta [100]. Efforts to integrate deep learning with PBMs has also extended to hydrological response modeling, flow estimation, and evapotranspiration simulation. These hybrid approaches have demonstrated superior fitting accuracy and generalization performance compared to traditional modeling techniques [100,101]. Collectively, existing studies demonstrate that integrated approaches not only surpass individual methods in terms of performance metrics but also significantly improve model adaptability and real-world applicability. The integration of RS, ML, and PBMs not only enhances water quality monitoring but also opens broad opportunities in addressing climate change and land use dynamics. These include applications in agricultural resource management, land surface and atmospheric studies, hydrological forecasting, and remote sensing image processing, underscoring its transformative potential across multidisciplinary domains of environmental and resource management.

6. Summary and Perspectives

RS, ML, and PBMs are the three pivotal technologies in contemporary water quality monitoring and prediction. Each offers distinct strengths such as spatial coverage, pattern recognition, and mechanistic understanding, while also exhibiting specific limitations that constrain their standalone effectiveness. RS has achieved remarkable success in large-scale perception and high-temporal resolution monitoring, particularly excelling in the inversion of optically active water quality parameters. The emergence of ML has enabled advanced data-driven interpretation of RS observations and timeseries forecasting, markedly improving model adaptability, predictive accuracy, and generalizability across diverse water bodies. In parallel, PBMs offer strong theoretical grounding and policy-relevant capabilities for simulating pollutant transport, transformation processes, and the long-term trends of state variables. This review highlights that integrating RS, ML, and PBMs into hybrid monitoring and modeling frameworks presents a promising strategy for addressing the complexity and uncertainty inherent in aquatic environmental systems.

Although the integration of RS, ML and PBMs is still in its early stages and not yet widely adopted for water quality monitoring and prediction, it has already demonstrated considerable scientific value and application potential. With continued advancements in algorithmic development and the growing availability of high-quality, multi-scale environmental observation data, this integrative paradigm is poised to become a foundational approach for future water quality assessment and the management of aquatic environments. Looking ahead, the integration of RS, ML, and PBMs is expected to advance along several key directions:

(1)

Standardization and Integration of Multi-Source Data: Efforts will focus on the standardization and efficient fusion of multi-source, heterogeneous datasets to build comprehensive spatiotemporal water environment databases.

(2)

Deep Learning–Mechanism Integration: The coupling of deep learning with physical process modeling will be further developed to enhance model interpretability and strengthen decision-support capabilities.

(3)

Lightweight, High-Performance Model Structures: The design of streamlined, high-efficiency models will enable deployment on edge-computing platforms and support real-time monitoring and early warning applications.

(4)

Evaluation and Adaptation Frameworks: A robust evaluation and adaptation system will be established to facilitate the cross-regional and cross-seasonal scalability of integrated models.

With the continuous advancement of RS technologies toward higher spatial, temporal, and spectral resolutions, the rapid evolution of intelligent algorithms, and the increasing refinement of PBMs, integrated modeling approaches are poised to become indispensable tools for global water quality monitoring, simulation, and management in aquatic ecosystems.