Nutrient and Dissolved Oxygen (DO) Estimation Using Remote Sensing Techniques: A Literature Review

Highlights

What are the main findings?

• There are two main model categories for predicting nutrient and DO concentration: empirical and artificial intelligence (AI). Model accuracy is strongly influenced by the size of the training dataset. Comparative studies conducted in the same study area have shown that ML and NNs achieve higher prediction accuracy than empirical models under the same sample size. However, the exact number of samples needed for either approach is not clearly provided in the literature. ML models (e.g., SVR, XGBoost) often outperform NNs when training data are limited, as they are generally less prone to overfitting under small-sample conditions.
• Current models are site-specific, and their transferability across different regions may be restricted. To improve their transferability, models should be validated across diverse regions and time periods, including a wide range of environmental conditions (e.g., trophic state, seasonality, depth and bottom characteristics).

What is the implication of the main finding?

• In Case II water bodies, eutrophication has emerged as a major water quality issue, calling for prompt management actions. Remote sensing facilitates its monitoring through changes in ocean color linked to chlorophyll a (chl a) fluctuations. While chl a indicates the effects of eutrophication, nutrient levels—its primary drivers—are key to predicting phytoplankton growth and implementing preventive measures.
• Long-term monitoring of nutrients and DO, combined with multiple water quality indicators (e.g., chl a) based on remotely sensed data, could enable more efficient assessment of the trophic state of water bodies and facilitate timely management actions.

Abstract

Eutrophication has emerged as a critical threat to water quality degradation and ecosystem health on a global scale, calling for prompt management actions. Remote sensing enables the monitoring of eutrophication by detected changes in ocean color caused by fluctuations in chlorophyll a (chl a). Although chl a is a crucial indicator of phytoplankton biomass and nutrient overloading, it reflects the outcome of eutrophication rather than its cause. Nutrients, the primary “drivers” of eutrophication, are essential indicators for predicting the potential phytoplankton growth in water bodies, allowing adoption of effective preventive measures. Long-term monitoring of nutrients combined with multiple water quality indicators using remotely sensed data could lead to a more precise assessment of the trophic state. Retrieving non-optically active constituents, such as nutrients and DO, remains challenging due to their weak optical characteristics and low signal-to-noise ratios. This work is an attempt to review the current progress in the retrieval of un-ionized ammonia (NH₃), ammonium (NH₄⁺), ammoniacal nitrogen (AN), nitrite (NO₂⁻), nitrate (NO₃⁻), dissolved inorganic nitrogen (DIN), phosphate (PO₄³⁻), dissolved inorganic phosphorus (DIP), silicate (SiO₂) and dissolved oxygen (DO) using remotely sensed data. Most studies refer to Case II highly nutrient-enriched water bodies. The commonly used spaceborne and airborne sensors, along with the selected spectral bands and band indices, per study area, are presented. There are two main model categories for predicting nutrient and DO concentration: empirical and artificial intelligence (AI). Comparative studies conducted in the same study area have shown that ML and NNs achieve higher prediction accuracy than empirical models under the same sample size. ML models often outperform NNs when training data are limited, as they are less prone to overfitting under small-sample conditions. The incorporation of a wider range of conditions (e.g., different trophic state, seasonality) into model training needs to be tested for model transferability.

1. Introduction

Over the past few decades, eutrophication has become one of the most widespread water quality issues globally. According to the European Marine Strategy Framework Directive (MSFD), eutrophication is defined as “a process driven by enrichment of water by nutrients, especially compounds of nitrogen (N) and/or phosphorus (P), leading to increased growth, primary production and biomass of algae, changes in the balance of organisms and poor water quality” [1]. Biodiversity loss and ecosystem degradation are recognized as major consequences of eutrophication, both in marine and freshwater bodies [2], affecting many sectors of local economies that rely on these natural resources (e.g., fisheries, recreation and drinking water quality) [3].

To determine the trophic state of water bodies, various water quality indicators are employed. Water quality indicators include physical, chemical and biological parameters that are quantitatively measured to present and communicate complex phenomena, such as trends and changes over time [4]. Nutrients, chlorophyll a (chl a), dissolved oxygen (DO), biochemical oxygen demand (BOD), chemical oxygen demand (COD), chemical oxygen demand using permanganate (CODMn) and Secchi disk depth (SDD) are the most widely utilized tropic state assessment indices [5,6,7].

Water quality indicators are traditionally determined by field measurements, sampling and laboratory analysis. Although conventional methods provide highly accurate results, they are costly and time-consuming, providing only point measurements. It is almost impossible to record the spatio-temporal variations in water quality characteristics on a regular basis [8]. This variability could be accomplished with the use of Geographic Information System (GIS), applying spatial interpolation techniques on in situ measurements for the estimation of various physicochemical and biological parameters. The GIS has also been applied in sampling strategies [9,10].

Since the early 1970s, remote sensing techniques have been widely integrated into the study of ocean color. Ocean color refers to the shade or tone of water which is determined by the interactions of electromagnetic radiation with the water itself and its constituents, including molecules and suspended and dissolved matter, both organic and inorganic [11]. It is typically defined as “the intensity and spectral distribution of visible light interacted with the water” [12], and measured variations in satellite imagery provide useful information about water quality indicators [8].

Ocean color remote sensing has been primarily focused on the retrieval of chl a, which is the main pigment found in phytoplankton. Monitoring chl a is crucial, as it serves as an indicator of total phytoplankton biomass and reflects the effects of nutrient overloading in marine and freshwater bodies. However, this condition signals the outcome of eutrophication, rather than its underlying cause. In contrast to chl a, nutrients, the primary “drivers” of eutrophication, could serve as crucial indicators for understanding the loading mechanisms and predicting phytoplankton growth potential in water bodies. Long-term monitoring of nutrients, combined with multiple water quality indicators (e.g., chl a and DO), using remotely sensed data could result in a more efficient assessment of the trophic state and prompt management actions.

Total nitrogen (TN) and total phosphorus (TP) represent the total nutrient load in marine and freshwater bodies and therefore are among the most commonly studied optically inactive water quality parameters (WQPs) in remote sensing-based assessments of eutrophication. Given that several existing review papers have covered TN and TP [8,13,14,15,16], these indicators are beyond the scope of this review.

This literature review aims to summarize the progress and the state of the art in retrieving non-optically active constituents of water such as un-ionized ammonia (NH₃), ammonium (NH₄⁺), ammoniacal nitrogen (AN), nitrite (NO₂⁻), nitrate (NO₃⁻), dissolved inorganic nitrogen (DIN), phosphate (PO₄³⁻), dissolved inorganic phosphorus (DIP), silicate (SiO₂) and dissolved oxygen (DO) using remotely sensed data. Commonly used spaceborne and airborne sensors are mentioned, along with selected spectral bands and derived band indices. Remote sensing methods for the quantitative retrieval of the selected WQPs, as well as their predictive accuracy based on various statistical indices, are also discussed. In addition, there is reference to the potential influence of depth, bottom characteristics, seasonality, trophic state and different concentration levels of nutrients and DO on the accuracy of the retrieval models.

2. Methodology

A systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology [17,18]. The literature search was performed using Scopus, selected for its broad, multidisciplinary coverage, including studies published up to June 2025. Eligible studies comprised peer-reviewed articles and conference papers that employed remote sensing data to estimate nutrients and DO. Reference lists of the selected articles were also checked for additional sources. Studies that focused exclusively on TN and TP were excluded, as they were already covered in prior reviews. Non-English studies, as well as studies unrelated to water quality (e.g., those focused on urban areas, land, crops, soil, groundwater, or agriculture), were also excluded. To ensure comprehensive coverage of relevant studies, the following queries were used in the topic field:

• (“nutrient”) AND (“remote sensing”) AND (“water quality”) AND (“empirical” OR “analytical” OR “semi-empirical” OR “artificial intelligence” OR “machine learning” OR “deep learning”) AND NOT (“total nitrogen” OR “TN” OR “total phosphorus” OR “TP” OR “urban” OR “land” OR “crop” OR “soil” OR “groundwater” OR “agriculture”);
• (“dissolved oxygen”) AND (“remote sensing”) AND (“water quality”) AND (“empirical” OR “analytical” OR “semi-empirical” OR “artificial intelligence” OR “machine learning” OR “deep learning”) AND NOT (“total nitrogen” OR “TN” OR “total phosphorus” OR “TP” OR “urban” OR “land” OR “crop” OR “soil” OR “groundwater” OR “agriculture”).

A total of (n = 1147) records were retrieved from the initial search: (n = 1136) records from electronic database (Scopus) and an additional (n = 11) records through reference lists of included articles. After removing duplicates (n = 102), the titles and abstracts of the remaining records (n = 1.045) were screened to exclude irrelevant studies, resulting in the removal of (n = 852). This process resulted in (n = 193) studies being retained for full-text assessment. Studies that focused exclusively on TN and TP were excluded (n = 74) at the full-text stage as not related to the aim of this study. Non-English studies (n = 36), as well as those unrelated to water quality parameters of interest (e.g., those that focused on urban areas, land, crops, soil, groundwater or agriculture) (n = 83), were also excluded. Finally, (n = 66) studies from 2001 to June 2025 were included in the systematic review (Figure 1), the majority of which referred to DO estimation (Figure 2). An in-depth analysis of the studies selected was then conducted. Additional information regarding the research strategy, PRISMA 2020 checklist (Table S1) and PRISMA flow diagram (Figure S1), is available in the Supplementary Materials.

For each included study, the following data were extracted:

• Bibliographic details (authors and year of publication);
• Parameters retrieved (NH₃, NH₄⁺, AN, NO₂⁻, NO₃⁻, PO₄³⁻, SiO₂, DIN, DIP or DO);
• Water type (coastal or inland; lake, river, reservoir, wetland, stream/stream network, lagoon or other);
• Location of the study area (country);
• Sensor (spaceborne, airborne, ground-based/multispectral or hyperspectral);
• Number of samples;
• Bands/band indices or equation;
• Retrieval model (empirical, machine learning (ML) or neural networks (NNs));
• Validation metrics (e.g., coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), etc.).

For studies that reported multiple sensors, all relevant results were extracted. In cases where multiple models were presented, the best performing model was recorded. Missing information or unreported values were recorded as “not available (NA)”. Due to methodological heterogeneity among studies, such as differences in sensor type, sample size and validation metrics, quantitative pooling (meta-analysis) was not feasible. Therefore, a narrative synthesis was conducted. Studies were grouped by retrieval model to identify patterns in retrieval accuracy based on the reported validation metrics.

3. Nutrients—The Primary “Drivers” of Eutrophication

N and P are essential nutrients for plant growth, survival and reproduction. N is typically the limiting factor in marine water bodies, while it is P in freshwater bodies. The limiting factor refers to the nutrient that is most crucial for the growth of primary producers, as it is required in large amounts and is in short supply [3,19]. N and P are found in both dissolved and particulate form, as well as in organic and inorganic form [20]. DIN and DIP refer to the dissolved forms of N and P and constitute immediately bioavailable molecules for primary producers. DIN includes NH₄⁺, NO₂⁻ and NO₃⁻, while DIP refers to PO₄³⁻ [19].

3.1. Un-Ionized Ammonia (NH₃), Ammonium (NH₄⁺) and Ammoniacal Nitrogen (AN)

In aquatic environments, ammonia consists of NH₃ and NH₄⁺ and depends on temperature and pH values. NH₄⁺ is the predominant form of ammonia in the pH range of the majority of natural water bodies (pH < 8.75). In contrast to NH₄⁺, NH₃ accounts for a relatively small fraction of ammonia and increases at a higher pH level and temperature. Although NH₃ is the most toxic form of ammonia, high concentrations of NH₄⁺ are toxic too and could affect the welfare of aquatic organisms. High levels of NH₄⁺ could lead to eutrophication, as it is the most preferred N source for phytoplankton growth. AN refers to the portion of N in the form of NH₃ and NH₄⁺ [21].

3.2. Nitrite (NO₂⁻) and Nitrate (NO₃⁻)

NH₄⁺ tends to be oxidized to NO₃⁻ in a two-step process (NH₄⁺ → NO₂⁻ → NO₃⁻) by aerobic chemoautotrophic bacteria, even if DO is found in low concentrations (1.0 mg/L) [22]. During this process, Nitrosomonas oxidize NH₄⁺ to NO₂⁻ and Nitrobacter oxidize NO₂⁻ to NO₃⁻. NO₂⁻ is an intermediate form of inorganic N, usually found in low concentrations, as it is typically converted to NO₃⁻ [23]. NO₃⁻ is generally present at higher concentrations than NH₄⁺ and NO₂⁻ [22]. Aquatic plants assimilate NH₄⁺ and NO₃⁻ and then N returns to water via the decomposition of dead organic matter [23]. NO₃⁻ is less toxic than NH₄⁺ and NO₂⁻ [22].

3.3. Phosphate (PO₄³⁻)

PO₄³⁻ is the only form of P that primary producers can assimilate. In contrast to other nutrients (e.g., NO₃⁻), PO₄³⁻ is not water-soluble and therefore enters aquatic environments only with soil movement, as it adheres to soil particles [3]. According to Kim et al. [24], PO₄³⁻ has no toxic effects on aquatic organisms.

3.4. Silicate (SiO₂)

Although most studies deal with the effects of eutrophication caused by N and P enrichment, silicon (Si), in the form of SiO₂, also, plays a crucial role in algal composition, controlling its growth. In aquatic environment, SiO₂ is an essential element for diatoms, a silica-shelled phytoplankton species. As a result, their growth rates are determined by the SiO₂ supply. SiO₂ enters coastal waters through rivers and therefore its concentration depends on this input load in particular [25]. In natural water bodies, toxicity from SiO₂ is not likely to occur [26].

3.5. Sources of Nutrient Enrichment

Eutrophication can occur naturally or be driven by human activities. The main natural sources are atmospheric deposition, river runoff, upwelling and submarine groundwater seepage (SGS) [19]. Atmospheric deposition refers to pollutants and other substances that are carried by wetfall (rain, snow, cloud and fog droplets) and dryfall (aerosols and particulates) to the Earth’s surface [23], while river runoff refers to freshwater that becomes nutrient-enriched by flowing through mineral-loaded rocks. Upwelling describes the “rising of cold waters masses from the deep that are enriched with nutrients, including NO₃⁻ and Si(OH)₄ that originate from organic matter rematerialized by bacteria on the seafloor” [19]. In other words, it is a mechanism that transports these nutrient-rich deep waters towards the surface. In addition to upwelling, winter mixing and eddies are able to transport significant amounts of nutrients to the water surface [19]. SGS refers to “the continental groundwater that is discharged directly into the ocean, wherever a coastal aquifer is connected to sea” [27]. SGS is recognized as an important transport pathway for a variety of substances, such as nutrients, nearshores, with significant environmental consequences [28]. Upwelling and SGS are primarily associated with marine water bodies, while atmospheric deposition and river runoff affect both marine and freshwater bodies.

Agricultural, industrial and urban wastewater runoffs are the primary anthropogenic sources contributing to increased primary production due to nutrient release [4]. Fertilizers rich in NO₃⁻, NH₄⁺ and PO₄³⁻ are flushed into rivers and groundwater through irrigation systems, rainfall and floods, eventually reaching marine water bodies. In addition, polluted water originating from domestic and industrial sources enriches marine and freshwater bodies with significant amounts of inorganic nutrients, as sewage treatment is rarely effective in removing N and P in forms of NH₄⁺ and PO₄³⁻ [19]. Aquaculture is another anthropogenic source of nutrient enrichment caused by the metabolic processes of cultivated organisms and uneaten food [29]. Although it contributes only 1% to nutrient release compared to other anthropogenic activities [30], its impact should not be underestimated.

3.6. Impacts on Aquatic Environment

Algal blooms, oxygen deficiency of bottom water, ecosystem degradation and loss of biodiversity are the main consequences of eutrophication [2]. Algal blooms refer to the proliferation and dominance of specific phytoplankton or macroalgal species that benefit from the rapid nutrients uptake rates. In marine and freshwater bodies, algal blooms can occur naturally. However, anthropogenic activities can significantly increase the intensity and frequency of this ecological process. Harmful substances produced by toxic algal blooms accumulate in filter feeders (e.g., shellfish) and, through them, be transported to higher levels of the trophic chain, which could lead to mortality of upper consumers (e.g., fish and humans).

Non-toxic blooms (e.g., red tides) produce high amounts of biomass in water columns, reducing the transmission of light and thus DO availability. The subsequent microbial decomposition of this biomass can lead to hypoxic or even anoxic conditions (DO < 2 mg/L) in deep waters. This DO reduction changes the redox potential of sediment and increase fluxes of PO₄³⁻, NH₄⁺ and SiO₂ into the overlaying water. Furthermore, bacterial sulfate respiration could replace DO, generating hydrogen sulfide (H₂S), a chemical compound toxic to most benthic macrofauna, leading to changes in species composition and loss of habitats and biodiversity [19].

4. Water Quality Monitoring Using Remote Sensing

4.1. Light–Water Interaction

When sunlight or another light source is directed towards the water surface, a portion of the light is reflected off the surface, while the remainder is transmitted through the water column. Underwater light is either absorbed or scattered by water molecules, particulate and dissolved matter. Only a little part of the incident light is scattered backwards and leaves the water. The back-scattered light gives water its characteristic color and corresponds to the water-leaving radiance detected by remote sensors [31].

Ocean color is determined by the absorption and scattering characteristics of water and its constituents. The processes of absorption and scattering refer to the inherent optical properties (IOPs) and define the intensity and spectral characteristics of water-leaving radiance [32]. IOPs are unique to each constituent [33]. Variations in water-leaving radiance are utilized to derive information about the types of constituents and their concentrations [34].

Remote sensing reflectance (R_rs) is a fundamental apparent optical property (AOP) of water commonly used to quantify ocean color. R_rs depends on the IOPs of the medium and the ambient field light and refers to the ratio of water-leaving radiance to downwelling irradiance, measured just above the water surface [12,31]. In other words, it represents the spectral distribution of the incident light that penetrates the water surface and returns through the air–water interface due to the differential absorption and scattering mechanisms of the water’s constituents. Understanding the optical properties of water and its constituents and hence the relationship between IOPs and the water-leaving radiance is essential for retrieving WQPs using remote sensing [12].

4.2. Optically Active and Inactive Constituents of Water

In ocean color remote sensing, WQPs are classified as optically active or inactive. Phytoplankton, yellow substances, also known as colored dissolved organic matter (CDOM) and inorganic particulate matter are the primary optically active constituents (OACs) that interact with light through absorption and scattering processes, giving water its characteristic color [33,35]. OACs can be successfully detected by remote sensors [34,35]. Chl a, the main phytoplankton pigment, absorbs the blue and red wavelengths of the visible spectrum and reflects green. Therefore, water bodies characterized by high phytoplankton abundance appear in shades of blue-green [31]. CDOM absorbs light in the ultraviolet (UV) and blue regions of the electromagnetic spectrum, giving water a yellow-green to brown tone [12]. A high concentration of inorganic particulate matter gives water a reddish-brown color, as it scatters longer wavelengths (red). On the other hand, water bodies characterized by low concentration of OACs absorb longer-wavelength red light and thus appear blue [31].

Water bodies are classified into Case I and Case II depending on whether their optical properties are dominated solely by phytoplankton or by multiple OACs. Case I refers to optically simple, deep water (open ocean), where the optical properties are primarily determined by phytoplankton. Case II refers to optically complex waters (coastal and freshwater bodies), where ocean color is significantly influenced by several constituents besides phytoplankton, which vary independently of each other (e.g., CDOM and inorganic particulate matter), and often they are accompanied by relatively high levels of scattering. The non-linear nature of Case II waters, along with the similarity in optical properties of various constituents, contributes to their optical complexity [12,34].

Non-optically active constituents, such as nutrients and DO, do not have a well-defined effect on the color of water [16,36], and they appear to exceed weak optical characteristics and low signal-to-noise ratios [8,16]. Several attempts have been made to estimate non-optically active constituents of water by establishing relationships with optically active ones. However, finding significant correlations between optically active and inactive WQPs remains a major challenge and may vary from one study to another [36,37].

4.3. Remote Sensing Imagery

To date, a variety of different remote sensing systems has been used to detect variations in ocean color by measuring the amount of radiation in different wavelengths reflected from the water’s surface [8]. Remote sensed data are available in digital format (images) captured by remote sensing sensors. The choice of a specific sensor depends on its spectral, spatial, temporal and radiometric resolution [16,33,38,39,40].

4.3.1. Optical Remote Sensing

Optical sensors have been widely used in the retrieval of WQPs, collecting data in the visible (VIS) and near-infrared (NIR) regions of the electromagnetic spectrum. Short-wave infrared (SWIR) and thermal infrared (TIR) regions are not directly related to ocean color; however, they are occasionally used for atmospheric correction and sea surface temperature (SST) detection, respectively. Optical sensors usually operate passively, relying on sunlight as the main source of illumination, and can be either spaceborne or airborne, depending on the platforms they are installed on. Spaceborne sensors are mounted on platforms flown outside the Earth’s atmosphere (satellites or aircraft), while airborne sensors are those carried by platforms to areas within the Earth’s atmosphere (unmanned aerial vehicles (UAVs), aircraft, helicopters, balloons and boats). Spaceborne and airborne sensors provide either multispectral or hyperspectral images, depending on the sensor used for image acquisition [16,33].

4.3.2. Microwave Remote Sensing

Microwave radiometers (MWR) and synthetic aperture radar (SAR) operate in the microwave portion of the electromagnetic spectrum. The longer wavelengths used by MWR and SAR, compared to VIS and NIR sensors, enable them to detect electromagnetic radiation under all weather conditions, except during the heaviest rains [16,41]. MWR are passive, while SAR systems operate actively by emitting their own energy. Although MWR and SAR are not directly related to ocean color and water quality monitoring, they provide complimentary data to optical ocean color observations. In particular, MWR provide datasets on SST, sea surface salinity (SSS), wind speed, water vapor, cloud liquid water, rain rate and sea ice [42], while SAR is able to derive data on ocean surface winds, ocean surface currents, sea level, ocean internal waves, sea ice, icebergs, aquatic vegetation, coral reefs, ocean oil spills, ocean tides, eddies and bathymetry [43]. Datasets from MWR and SAR improve the accuracy of ocean color data by accounting for atmospheric conditions and oceanographic processes that influence variations in ocean color.

4.4. Remotely Sensed WQP Retrieval Methods

In general, four different approaches can be used for the retrieval of WQPs through remote sensing: the empirical, analytical, semi-empirical and AI methods [15].

4.4.1. Empirical Method

The empirical method determines statistical relationships between in situ water quality measurements and AOPs of water. The spectral band and/or band combinations that have the highest correlation with WQPs are identified and selected as the remotely sensed data. Then, the statistical relationship between the in situ measurements and reflectance data is determined using different linear or non-linear regression techniques [44]. The empirical method is applied in the retrieval of both optically active and inactive WQPs [13,15].

4.4.2. Analytical Method

The analytical method is based on the light–water interaction, utilizing the radiative transfer equation which defines the relationship between the IOPs and the AOPs of water [12]. The concentration of water constituents can be estimated using a ratio of their absorption and backscattering coefficients [36,44]. Optically active WQPs can be successfully derived using the analytical method [13,36].

4.4.3. Semi-Empirical Method

The semi-empirical approach is a combination of the analytical and the empirical methods, as it integrates the radiative transfer theory with traditional statistical techniques. The IOPs of water are considered to identify appropriate spectral bands and/or band combinations and regression techniques are applied to establish relationships for their retrieval [44]. Semi-empirical models are, in particular, used for the estimation of optically active WQPs [13,45].

4.4.4. AI Method

AI refers to “any technique that aims to enable computers to mimic human behavior and reproduce over human decision-making to solve complex tasks independently or with minimum human intervention” [46]. In recent years, AI methods, such as ML and NNs, have been increasingly applied in various real-world application areas [47], including, among others, ocean color remote sensing [15,16]. ML and NNs aim to identify new insights, patterns, relationships and dependencies in data, make predictions and assist in data-driven decision-making [47].

ML

ML is an interdisciplinary field under AI that combines both statistics and computer science [48]. According to [46], ML aims to “automate the task of analytical model building by applying algorithms that iteratively learn from problem-specific training data, allowing computers to discover hidden insights and complex patterns without being explicitly programmed”. Based on the given problem and the available input data, algorithms can be grouped into three ML techniques:

• Supervised learning;
• Unsupervised learning;
• Reinforcement learning.

In supervised learning, classification and regression are the main methods used. In classification, target prediction (output) is categorical or discrete, while in regression, it is numerical or continuous. In both the classification and regression methods, the building of the predictive model is based on pre-existing data and certain algorithms can be applied, depending on the desired output. On the other hand, in unsupervised learning, there are no pre-existing data labels to train predictive models. Clustering and association are the main methods used in unsupervised learning. In clustering, each data entry is aligned in groups (clusters) that share similar characteristics, while in association, the dependency between one data entry to another is found to map them together [48].

NNs

NNs are a subfield of ML which refer to “a collection of connected units (neurons) organized in layers”. The architecture of NNs consists of three types of layers, including the input, the hidden (middle) and the output layer. Based on the number of hidden layers, NNs are divided into shallow (a single hidden layer) and deep (multiple hidden layers). A larger number of hidden layers enables deep NNs to learn high-level features with more complexity than shallow NNs [49].

5. Narrative Literature Synthesis

Because of substantial heterogeneity among studies, including differences in study area, sensor type, sample size and validation metrics, a quantitative synthesis (meta-analysis) was not feasible. Therefore, a narrative synthesis was conducted. Studies were grouped by WQP, study area (inland or coastal waters) and retrieval model to identify patterns in retrieval accuracy based on the reported validation metrics. The estimation of nutrients and DO relied on proxy relationships, rather than on the direct optical sensitivity of the parameters, as these constituents are non-optically active. A summary of studies that have retrieved nutrients and DO using remotely sensed data is provided in

Table 1. Remote sensing-based retrieval of nutrients and DO in inland waterbodies using empirical models.

WQP	Study Area	Sensor	N	Model	Equation or Band/Band Indices	Accuracy	Reference
NH3	Guanting Reservoir (CN)	Landsat 5 (TM)	76	MLR stepwise	$=-7.177+1.93\mathrm{l}\mathrm{n}\mathrm{B}7+0.1323\mathrm{B}6-2.185\mathrm{B}6/\mathrm{B}3-0.07648\mathrm{B}1$	Sig. = 0.00, R = 0.81, MRE(%) = 28.00%	[50]
	Danjiangkou Reservoir (CN)	Sentinel-2 (MSI)	140	SLR	$=0.296{\mathrm{B}3/\mathrm{B}2}^2-0.224\mathrm{B}3/\mathrm{B}2-0.328$	R2 = 0.85, MAE = 0.01, RMSE = 0.01	[51]
NH4+	Karla Lake (GR)	Landsat 7 (ETM+)	NA	SLR	$\mathrm{B}2/\mathrm{B}4$	R2(%) = 94.32%	[52]
		Landsat 8 (OLI)			$(\mathrm{B}1-\mathrm{B}3)/\mathrm{B}2$	R2(%) = 80.64%
	Xiangxi River (CN)	HJ-1	NA	NLR	$=10.365{\mathrm{B}4/(\mathrm{B}2\times \mathrm{B}3)}^{0.7042}$	R2 = 0.34	[53]
	Trichonis Lake (GR)	Landsat 8 (OLI)	44	SLR	$=-0.323+0.136\times \left[\left(\mathrm{B}1-\mathrm{B}4\right)/\left(\mathrm{B}3-\mathrm{B}4\right)\right]$	R2 = 0.47, R = 0.69, SE = 0.00, F value = 5.42, Sig. = 0.01	[54,55]
	Erlong Lake (CN)	Landsat 5 (TM), Landsat 7 (ETM+), Landsat 8 (OLI)	31	SLR	$=0.8\times \left(\left(\mathrm{R}\mathrm{e}\mathrm{d}+\mathrm{N}\mathrm{I}\mathrm{R}\right)+\left({\displaystyle \frac{\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}}{\mathrm{N}\mathrm{I}\mathrm{R}}}\right)\times \left(\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}+\mathrm{N}\mathrm{I}\mathrm{R}\right)+\left({\displaystyle \frac{\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}}{\mathrm{N}\mathrm{I}\mathrm{R}}}\right)\right)\times 0.099$	R2 = 0.86, RMSE = 0.65, Sig. =0.00 1	[56]
						R2 = 0.95, RMSE = 0.53 2
NO2−	Mosul Dam Lake (IQ)	Landsat 5 (TM)	12	SLR	$=0.109-0.003\mathrm{B}3$	R2 = 0.92, SEE = 0.00, Sig. = 0.01	[57]
		Landsat 7 (ETM+)		MLR	$=-0.658+0.008\mathrm{B}4+0.012(\mathrm{B}1/\mathrm{B}3)+0.569(\mathrm{B}62/\mathrm{B}61)$	R2 = 0.99, SEE = 0.01, Sig. = 0.00
NO3−	Porsuk Dam Reservoir (TR)	Landsat 4 (TM)	NA	MLR Stepwise	$=2.84-0.06\mathrm{B}1-0.05\mathrm{B}2+0.06\mathrm{B}3+0.38\mathrm{B}4$	R2 = 0.86	[58]
	Guanting Reservoir (CN)	Landsat 5 (TM)	76	MLR Stepwise	$=21.9888-0.70578\mathrm{B}2/\mathrm{B}5-3.688\mathrm{l}\mathrm{n}\mathrm{B}6-0.0186\mathrm{B}3$	Sig. = 0.00, R = 0.92, MRE(%) = 4.40%	[50]
	Wisconsin streams (USA)	MODIS	315	xPLSR	BGT, SOI, EVI, DST, NPV, NDV, GRN, GVF, LAI, FPR, WET, GPP	R2 = 0.80, Bias = 0.00	[59]
	Mosul Dam Lake (IQ)	Landsat 5 (TM)	12	NLR	$=0.2394\times {\left(\mathrm{B}2/\mathrm{B}3\right)}^{6.7012}$	R2 = 0.93, SEE = 0.10, Sig. = 0.01	[57]
		Landsat 7 (ETM+)		SLR	$=1.782+75.469\ln (\mathrm{B}62/\mathrm{B}61)$	R2 = 0.60, SEE = 1.51, Sig. = 0.04
	Karla Lake (GR)	Landsat 7 (ETM+)	NA	SLR	$\mathrm{B}4/\mathrm{B}7$	R2(%) = 55.50%	[52]
		Landsat 8 (OLI)			$\mathrm{l}\mathrm{n}\mathrm{B}6-\mathrm{l}\mathrm{n}\mathrm{B}7$	R2(%) = 55.50%
	Xiangxi River (CN)	HJ-1	NA	SLR	$=-663.72{\mathrm{B}1/(\mathrm{B}2+\mathrm{B}4)}^3+800.45{\mathrm{B}1/(\mathrm{B}2+\mathrm{B}4)}^2-309.81\mathrm{B}1/\left(\mathrm{B}2+\mathrm{B}4\right)+39.18$	R2 = 0.75	[53]
	Dam Lake of Wadi Baysh (SA)	Sentinel-2 (MSI)	120	MLR	$(\mathrm{B}8-\mathrm{B}3)/(\mathrm{B}8+\mathrm{B}3)$	R2 = 0.94, Adj. R2 = 0.94, RMSE = 0.07, Mean Response = 0.49	[60]
NO3−	Bin El Ouidance Reservoir (MA)	Sentinel-2 (MSI)	19	MLR Stepwise	$=0.095\times \mathrm{B}1-0.014\times \mathrm{B}2-0.062\times \mathrm{B}3-0.0126\times \mathrm{B}9-66.442$	R2 = 0.67, RMSE = 0.62	[61]
	Erlong Lake (CN)	Landsat 5 (TM), Landsat 7 (ETM+), Landsat 8 (OLI)	31	SLR	$=\left(\mathrm{R}\mathrm{e}\mathrm{d}+\mathrm{N}\mathrm{I}\mathrm{R}\right)+{\mathrm{e}}^{\left({\displaystyle \frac{\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}}{\mathrm{N}\mathrm{I}\mathrm{R}}}\right)\times (0.029)}+{\mathrm{e}}^{{\displaystyle \frac{\frac{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}{\mathrm{N}\mathrm{I}\mathrm{R}}}{\mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}+\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}}\times \left(\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}\right)}\times (0.5)$	R2 = 0.88, RMSE = 8.50, Sig. = 0.00 1	[56]
						R2 = 0.99, RMSE = 1.05 2
	Guartinaja, Sapal, Momil Wetlands (CO)	Sentinel-2 (MSI)	NA	MLR	$=-88.3612+2119.4812\times \mathrm{B}6-2338.2696\times \mathrm{B}8+99.7217\times (\mathrm{B}8/\mathrm{B}6)$	R2 = 0.86, RMSE = 0.20	[62]
${\mathrm{P}\mathrm{O}}_4^{3-}$	Guanting Reservoir (CN)	Landsat 5 (TM)	76	MLR Stepwise	$=-0.0698+1.9\times {10}^{-5}\times {\mathrm{e}}^{\mathrm{B}3/10}+0.0855{\displaystyle \frac{\mathrm{B}3}{\mathrm{B}2}}$	Sig. = 0.00, R = 0.96, MRE(%) = 15%	[50]
	Winconsin streams (USA)	MODIS	315	xPLSR	EVI, SOI, DST, ROD, GRN, GVF, LAI, URB, ROI, NDV	R2 = 0.51, Bias = 0.00	[59]
	Rosetta River Nile Branch (EG)	Worldview-2	38	MLR Stepwise	$=-0.516+(81.002/\mathrm{B}7)-(131.951/\mathrm{B}5)$—East Side	R2 = 0.19	[63]
					$=0.239-(139.542/\mathrm{B}4)$—West Side	R2 = 0.80
${\mathrm{P}\mathrm{O}}_4^{3-}$	Mosul Dam Lake (IQ)	Landsat 5 (TM)	12	SLR	$=-3.783+0.264\mathrm{B}5$	R2 = 0.75, SEE = 0.48, Sig. = 0.05	[57]
		Landsat 7 (ETM+)		MLR	$=-0.081-0.008\mathrm{B}3+0.018\mathrm{B}4$	R2 = 0.96, SEE = 0.03, Sig. = 0.00
	Xiangxi River (CN)	HJ-1	NA	SLR	$=-0.6677{(\mathrm{B}1-\mathrm{B}4)/(\mathrm{B}1+\mathrm{B}4)}^2+0.8232(\mathrm{B}1-\mathrm{B}4)/\left(\mathrm{B}1+\mathrm{B}4\right)-0.1494$	R2 = 0.18	[53]
	Wular Lake (IN)	Landsat 8 (OLI)	11	SLR	$=57.848{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{L}\mathrm{P}}+5.0929$	R2 = 0.73, Sig. = 0.00	[64]
	Quaroum Lake (EG)	ASTER	18	MLR	$=0.0006({\mathrm{B}5)}^2-0.0388\left(\mathrm{B}5\right)+0.7684$	R2 = 0.94, RMSE = 0.01, SEE = 0.01, Sig. = 0.01 1	[65]
						RMSE = 0.01, SEE = 0.00 2
	Bin El Ouidance Reservoir (MA)	Sentinel-2 (MSI)	19	MLR Stepwise	$=0.010\times \mathrm{B}1-0.006\times \mathrm{B}3-0.022\times \mathrm{B}4+0.0388\times \mathrm{B}5-0.0105\times \mathrm{B}8\mathrm{A}-0.0155\times \mathrm{B}10-8.507$	R2 = 0.54, RMSE = 1.02	[61]
SiO2	Xiangxi River (CN)	HJ-1	NA	SLR	$=-295.08{\mathrm{B}1/(\mathrm{B}2+\mathrm{B}4)}^2+268.58\mathrm{B}1/\left(\mathrm{B}2+\mathrm{B}4\right)-52.123$	R2 = 0.56	[53]
DIN	Burullus Lake (EG)	Landsat (TM)	NA	MLR Stepwise	NA	NA	[66]
DO	Mazala Lagoon (EG)	Landsat 5 (TM)	NA	MLR Stepwise	NA	NA	[67]
	Taihu Lake (CN)	Landsat 5 (TM)	15	NLR	$={\mathrm{e}}^{(2.3704-0.2107\times \ln \mathrm{B}3)}$	NA	[68]
	Burullus Lake (EG)	Landsat (TM)	NA	MLR	NA	NA	[66]
	Huangpu River (CN)	Landsat (TM)	NA	MLR	$=6.5088{\mathrm{e}}^{-0.2322\mathrm{B}4/\mathrm{B}2}$	R2 = 0.68	[69]
	Al-Saad Lake (SA)	Worldview-2	46	MLR	$=9.679+\left(\mathrm{B}6/\mathrm{B}8\times -3.038\right)+\left(\mathrm{B}4-\mathrm{B}7/\mathrm{B}4+\mathrm{B}8\times -4.793\right)+(\mathrm{B}4/\mathrm{B}8\times 0.752)$	R2 = 0.67	[70]
	Rosetta River Nile Branch (EG)	Worldview-2	38	MLR Stepwise	$=-3.706-(124.179/\mathrm{B}7)+(213.452/\mathrm{B}6)+(1033.090/\mathrm{B}3)$—East Side	R2 = 0.44	[63]
					$=0.555+(307.840/\mathrm{B}6)-(122.283/\mathrm{B}5)-(484.740/\mathrm{B}1$)—West side	R2 = 0.38
	Gomti River (IN)	IRS LISS III	NA	MLR	$=-6.747+1.057\times \mathrm{B}1+5.298\times (\mathrm{B}3/\mathrm{B}1)+9.146\times (\mathrm{B}4/\mathrm{B}1)$—Pre monsoon	R2 = 0.76	[71]
					$=-8.915+3.085\times \mathrm{B}1-1.674\times (\mathrm{B}3/\mathrm{B}1)+4.475\times (\mathrm{B}4/\mathrm{B}1)$—Post monsoon	R2 = 0.57
	Karla Lake (GR)	Landsat 7 (ETM+)	NA	SLR	$(\mathrm{B}2+\mathrm{B}3)/2$	R2(%) = 88.53%	[52]
		Landsat 8 (OLI)			$\mathrm{B}1/\mathrm{B}3$	R2(%) = 80.49%
	Bardawil Lagoon (EG)	Landsat 8 (OLI)	NA	MLR Stepwise	$=2.7749+1.67414\times \mathrm{B}2/\mathrm{B}4$—Spring	R2 = 0.57, SE = 0.32, RMSE = 0.39, Bias = 0.00	[72]
					$=-8.6889+3.59084\times \mathrm{B}2/\mathrm{B}3+13.73666\times \mathrm{B}7/\mathrm{B}6$—Summer	R2 = 0.78, SE = 0.23, RMSE = 0.24, Bias = 0.00
					$=2.2881+1.68452\times \mathrm{B}2/\mathrm{B}4$—Autumn	R2 = 0.33, SE = 0.26, RMSE = 0.42, Bias = 0.00
DO					$=-3.1377+6.51935\times \mathrm{B}2/\mathrm{B}3$—Winter	R2 = 0.67, SE = 0.31, RMSE = 0.35, Bias = 0.00
	Wular Lake (IN)	Landsat 8 (OLI)	11	SLR	$=355{\mathrm{P}\mathrm{C}}_4+6.325$	R2 = 0.43, Sig. = 0.03	[64]
	El Guajaro Reservoir (CO)	Landsat 8 (OLI)	NA	MLR Stepwise	$=37.182+223510000\times {\left(\mathrm{B}1\right)}^8+72725\times {\left(\mathrm{B}3\right)}^5-122280\times {\left(\mathrm{B}4\right)}^5-3.5878\times \left({\displaystyle \raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$\mathrm{B}5$}\right.}\right)-1325.5\times \left(\mathrm{B}7\right)+85.887\times \left({\displaystyle \raisebox{1ex}{$\mathrm{B}7$}\!\left/ \!\raisebox{-1ex}{$\mathrm{B}5$}\right.}\right)-1794000000\times {\left(\mathrm{B}3\right)}^{10}$	R2 = 0.93, RMSE = 0.09	[73]
	Tubay River (PH)	Landsat 8 (OLI)	NA	MLR Enter	$=7.961+65.873\times \mathrm{P}\mathrm{C}5{\_\mathrm{S}\mathrm{R}}_1-38.644\times {\mathrm{S}\mathrm{R}}_2\mathrm{B}1+58.405\times \mathrm{P}\mathrm{C}4{\_\mathrm{S}\mathrm{R}}_2+1.560\times {\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}}_2-1.378\times \mathrm{B}\mathrm{R}+9.507\times \mathrm{R}\mathrm{T}\mathrm{O}\mathrm{A}\mathrm{B}7$	R2(%) = 100%	[74]
				MLR Forward	$=7.446+0.008\times \mathrm{T}\mathrm{u}\mathrm{r}\mathrm{b}\mathrm{i}\mathrm{d}\mathrm{i}\mathrm{t}\mathrm{y}-11.59\times {\mathrm{S}\mathrm{R}}_2\mathrm{B}1$	R2(%) = 88.50%, SE = 0.83
	Bin El Ouidance Reservoir (MA)	Sentinel-2 (MSI)	19	MLR Stepwise	$=-0.0167\times \mathrm{B}8+0.0067\times \mathrm{B}9+0.0162\times \mathrm{B}10+0.0083\times \mathrm{B}11+9.577$	R2 = 0.74, RMSE = 0.20	[61]
	Três Marias Reservoir (BR)	Sentinel-2 (MSI)	13	MLR	$=9.2505+\left(-171.0251\times \mathrm{B}2\right)+\left(236.9708\times \mathrm{B}4\right)+\left(76.8288\times \mathrm{B}6\right)+\left(-150.7815\times \mathrm{B}11\right)$—30 × 30 m kernels	R2 = 0.85, MAE = 0.06, RMSE = 0.07, NRMSE = 36.2, p < 0.01	[75]
					$=8.9055+\left(-129.5866\times \mathrm{B}2\right)+\left(192.3651\times \mathrm{B}4\right)\times \left(36.4049\times \mathrm{B}6\right)+(-116.7094\times \mathrm{B}11)$—90 × 90 m kernels	R2 = 0.83, MAE = 0.06, RMSE = 0.08, NRMSE = 38.6, p < 0.01
		Landsat 8 (OLI)			$=9.5867+\left(-127.1909\times \mathrm{B}2\right)+\left(115.4625\times \mathrm{B}4\right)+\left(-223.5492\times \mathrm{B}6\right)+\left(227.0583\times \mathrm{B}7\right)$—90 × 90 m kernels	R2 = 0.69, MAE = 0.08, RMSE = 0.11,
DO					$=9.5867+\left(-127.1909\times \mathrm{B}2\right)+\left(115.4625\times \mathrm{B}4\right)+\left(-223.5492\times \mathrm{B}6\right)+\left(227.0583\times \mathrm{B}7\right)$—90 × 90 m kernels	NRMSE = 53.10, p = 0.03
	Erlong Lake (CN)	Landsat 5 (TM), Landsat 7 (ETM+), Landsat 8 (OLI)	31	SLR	$=\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}\right)\times \left(\left({\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}}\right)\times 80\right)+8.3$	R2 = 0.30, RMSE = 0.75, Sig. = 0.00 1	[56]
						R2 = 0.62, RMSE = 1.39 2
	Tigris River (IQ)	Landsat 8 (OLI)	NA	LASSO	NA	R2 = 0.76, RMSE = 0.25	[76]
	Guartinaja, Sapal, Momil Wetlands (CO)	Sentinel-2 (MSI)	NA	MLR	$=-39.2556+0.8061/\mathrm{B}4+4288.3263\times \left(\mathrm{B}4\times \mathrm{B}5\right)+19.4829\times (\mathrm{B}4/\mathrm{B}5)$	R2 = 0.95, RMSE = 0.18	[62]

¹ Training set, ² testing set.

,

Table 2. Remote sensing-based retrieval of nutrients and DO in coastal waterbodies using empirical models.

WQP	Study Area	Sensor	N	Model	Equation or Bands/Band Indices	Accuracy	Reference
NH3	Qinzhou Bay (CN)	DJI Elf 4 UAV	NA	PLSR	$=0.73\times \mathrm{B}1-0.71\times \mathrm{B}2+0.29\times \mathrm{B}3-1.41\times \mathrm{B}4+1.76\times \mathrm{B}5$	R2 = 0.56, RMSE = 1.06, MAE = 0.83 1	[77]
						R = 0.43, RMSE = 0.78, MAE = 0.60 2
NO2−	Qinzhou Bay (CN)	DJI Elf 4 UAV	NA	PLSR	$=-0.58\times \mathrm{B}1+0.12\times \mathrm{B}2+0.21\times \mathrm{B}3-0.16\times \mathrm{B}4+0.52\times \mathrm{B}5$	R2 = 0.33, RMSE = 0.52, MAE = 0.32 1	[77]
						R = 0.57, RMSE = 0.23, MAE = 0.18 2
NO3−	Qinzhou Bay (CN)	DJI Elf 4 UAV	NA	PLSR	$=-1.30\times \mathrm{B}1+0.92\times \mathrm{B}2+0.25\times \mathrm{B}3-0.63\times \mathrm{B}4+1.00\times \mathrm{B}5$	R2 = 0.59, RMSE = 0.74, MAE = 0.45 1
						R = 0.81, RMSE = 0.32, MAE = 0.25 2
${\mathrm{P}\mathrm{O}}_4^{3-}$	West Coast of Mumbai (IN)	IKONOS	NA	MLR	$=0.858\times \mathrm{B}1+0.209\times \mathrm{B}2+0.464\times \mathrm{B}3-0.188$—Coast	R2 = 0.98	[78]
					$=3.122\times \mathrm{B}1-7.97\times \mathrm{B}2-1.47\times \mathrm{B}3+0.812$—Creek
					$=-6.184\times \mathrm{B}1+49.06\times \mathrm{B}2-11.032\times \mathrm{B}3-3.95$—Seashore
		Landsat 7 (ETM+)	NA	MLR	$=0.323\times \mathrm{B}1+0.918\times \mathrm{B}2+0.887\times \mathrm{B}3-0.125$—Coast	R2 = 0.97
					$=-7.86\times \mathrm{B}1-13.38\times \mathrm{B}2-11.58\times \mathrm{B}3+2.83$—Creek
					$=2.96\times \mathrm{B}1+26.97\times \mathrm{B}2-4.08\times \mathrm{B}3-2.8$—Seashore
	Qinzhou Bay (CN)	DJI Elf 4 UAV	NA	PLSR	$=0.89\times \mathrm{B}1-1.42\times \mathrm{B}2+0.46\times \mathrm{B}3+0.33\times \mathrm{B}4-0.36\times \mathrm{B}5$	R2 = 0.48, RMSE = 0.46, MAE = 0.37 1	[77]
						R = 0.72, RMSE = 0.36, MAE = 0.29 2
DIN	Haizhou Bay (CN)	MODIS	41	SLR	$=7.86(\mathrm{B}3+\mathrm{B}4)/\left(\mathrm{B}3-\mathrm{B}4\right)+9.83$—70 μg/L ≥ DIN ≥ 70 μg/L	R2 = 0.72, AAE = 31.10, RMSE = 34.40 1	[79]
					$=7.86(\mathrm{B}3+\mathrm{B}4)/\left(\mathrm{B}3-\mathrm{B}4\right)+9.83$—70 μg/L ≥ DIN ≥ 70 μg/L	AAE = 35.50, RMSE = 37.20 2
					$=7.403(\mathrm{B}3+\mathrm{B}4)/\mathrm{B}3-\mathrm{B}4)+37.14$—DIN > 70 μg/L	R2 = 0.88, MAE = 18.22, RMSE(%) = 23.97% 1
						MAE = 13.50, RMSE(%) = 17.48% 2
	Bohai Sea (CN)	MODIS	51	MLR Stepwise	$=-4.514+0.011\times {\left(\right(\mathrm{B}8+\mathrm{B}10)/(\mathrm{B}8-\mathrm{B}10))}^3-0.159\times \left(\right({\mathrm{B}3+\mathrm{B}4)/(\mathrm{B}3-\mathrm{B}4))}^3+0.008\times {\left(\right(\mathrm{B}14+\mathrm{B}21)/(\mathrm{B}14-\mathrm{B}21))}^3-0.001\times \left(\right({\mathrm{B}12+\mathrm{B}15)/(\mathrm{B}12-\mathrm{B}15))}^3$— Bohai-Iaizhou Bay	R2 = 0.85, RMSE = 39.43, RPD = 2.58 1	[80]
						R2 = 0.82, RMSE = 41.12, RPD = 2.27 2
			10		$=30.960+0.0000009199\times {\left(\right(\mathrm{B}11+\mathrm{B}12)/(\mathrm{B}11-\mathrm{B}12))}^3$—Liadong Bay	R2 = 0.98, RMSE = 17.04, RPD = 7.10 1
						R2 = 0.99, RMSE = 7.18, RPD = 21.98 2
			59		$=13.532+0.160\times \left(\right({\mathrm{B}1+\mathrm{B}5)/(\mathrm{B}1-\mathrm{B}5))}^3+0.007\times {\left(\right(\mathrm{B}8+\mathrm{B}19/\mathrm{B}8-\mathrm{B}19\left)\right)}^3$—Inner Sea	R2 = 0.77, RMSE = 3.74, RPD = 2.07 1
						R2 = 0.79, RMSE = 2.35, RPD = 2.05 2
			120		$=-69.562+0.008\times \left(\right({\mathrm{B}8+\mathrm{B}10)/(\mathrm{B}8-\mathrm{B}10))}^3+0.044\times {\left(\right(\mathrm{B}1+\mathrm{B}20)/(\mathrm{B}1-\mathrm{B}20))}^3$—Entire Bohai Sea	R2 = 0.60, RMSE = 57.49, RPD = 1.58 1
						R2 = 0.68, RMSE = 49.23, RPD = 1.74 2
	Qinzhou Bay (CN)	DJI Elf 4 UAV	NA	PLSR	$=-1.14\times \mathrm{B}1+0.31\times \mathrm{B}2+0.75\times \mathrm{B}3-2.20\times \mathrm{B}4+3.29\times \mathrm{B}5$	R2 = 0.57, RMSE = 1.87, MAE = 1.17 1	[77]
						R = 0.76, RMSE = 0.85, MAE = 0.66 2
DO	Eastern coastal region of the Yellow Sea (KR)	MODIS, VIIRS	166	MLR Stepwise	$=-0.131\times \mathrm{S}\mathrm{S}\mathrm{T}-0.132\times {\mathrm{S}\mathrm{S}\mathrm{T}}_{\mathrm{m}-1}+0.066\times {\mathrm{c}\mathrm{h}\mathrm{l}a}_{\mathrm{m}-1}+12.343$	ARE(%) = 89.20%, IOA(%) = 78.60%	[81]
	Tangier—Ksar Sghir coastline (MA)	Landsat 8 (OLI)	NA	MLR	$=7.98+16.625\left(\mathrm{B}1\right)-23.839\left(\mathrm{B}4\right)$	R2 = 0.73, SEE = 0.22, p < 0.0001, RMSE = 0.25, Tol. = 0.24 1	[82]
						SEE = 0.47, RMSE = 0.69 2
	Qinzhou Bay (CN)	DJI Elf 4 UAV	NA	PLSR	$=1.60\times \mathrm{B}1+0.73\times \mathrm{B}2-2.27\times \mathrm{B}3+3.59\times \mathrm{B}4-3.79\times \mathrm{B}5$	R2 = 0.26, RMSE = 2.43, MAE = 1.72 1	[77]
						R = 0.51, RMSE = 1.27, MAE = 1.05 2
	Zhejiang coastal waters (CN)	Landsat 8 (OLI), Landsat 9 (OLI-2)	95	MLR	$=38.1165\times \mathrm{B}4+0.88\times (\mathrm{B}2/\mathrm{B}8)-0.069\times \mathrm{S}\mathrm{S}\mathrm{T}+7.226$	R2 = 0.59, Adj. R2 = 0.58 1	[83]
						R2 = 0.70, Adj. R2 = 0.64, RMSE = 0.55 2
	Coastal waters of the Gaza Strip (PS)	Sentinel-2 (MSI)	NA	MLR	$=-21.2535+\left(23.96945\times \left(\mathrm{B}1\right)\right)+\left(161.0584\times \left(\mathrm{B}2\right)\right)+(14.51334\times \left(\mathrm{B}3\right))$	R2 = 0.73, RMSE(%) = 0.21%, MAPE(%) = 6.60%	[84]

¹ Training set, ² testing set.

,

Table 3. Remote sensing-based retrieval of nutrients and DO in inland waterbodies using ML, shallow and deep NN models.

WQP	Study Area	Sensor	N	Model	Algorithm	Bands/Band Indices	Accuracy	Reference
NH3	Weihe River (CN)	SPOT 5	NA	ML	SS-SVR	B1, B2, B3, SWIR	R2 = 0.98, MSE = 0.05	[85]
	Weihe River (CN)	SPOT 5	13	ML	GA-SVR	B1, B2, B3, SWIR	R2 = 0.98, MSE = 0.77, MAD = 0.43	[86]
	Q Reservoir (CN)	Sentinel-2 (MSI)	NA	ML	XGBoost	B2, B3, B4, B5, B6, B7, B8, B8A	R2 = 0.82, RMSE = 0.09, MAPE = 28.60, Bias = −21.80	[87]
	Hulum Lake (CN)	Landsat 8 (OLI)	221	ML	RF	B1− B5, B4 − B5, B6/B7, B3 − B5, (B6 − B7)/(B6 + B7), B2 − B5, (B3 − B5)/(B3 + B5), B5 + B6	R2 = 0.71, MAE = 0.09, RMSE = 0.13	[88]
	Hongjianao Lake (CN)	Sentinel-2 (MSI)	NA	ML	RF-SHAP	B3 + B7	R2 = 0.59, RMSE = 0.11, MAE = 0.09, RPD = 1.58	[89]
	Nanfei River (CN)	Red edge-MX UAV	67	ML	GA-XGBoost	(B3 + B4)/B2, (B2 + B3 + B4)/B2, (B3 + B4)/(B1 + B2), (B2 + B3 + B4)/(B3 + B4)	R2 = 0.69, MAE = 0.14, RMSE = 0.16	[90]
	Yuandang Lake (CN)	DJI P4 UAV	60	ML	GB	B3−1 − B4−1	R2 = 0.84, RMSE = 0.06 1	[91]
							R2 = 0.55, RMSE = 0.12, MAE = 0.09 2
	Yuhe River (CN)	Ground- based	NA	ML	ABR	Red	R2 = 0.93, RMSE = 0.08, MAE = 0.07, MAPE(%) = 22.93%, Acc.(%) = 71.67% 1	[92]
							R2 = 0.95, RMSE = 0.12, MAE = 0.10, MAPE(%) = 22.25, Acc.(%) = 78.95% 2
NH3	Quinwu, Longjing, Nanshan Reservoirs (CN)	Sentinel-2 (MSI)	NA	NNs	MDN	B2, B3, B4, B5, B6, B7, B8	NA	[93]
		Landsat 7 (ETM+)				B1, B2, B3, B4
		DJI M300 UAV				B1, B2, B3, B4, B5, B6, B7, B8, B9, B10	R2 = 0.78, RMSE = 0.17, MAE = 0.04 1
							R2 = 0.98, RMSE = 0.09, MAE = 0.04 2
	Guanhe River (CN)	Airborne	60	NNs	Patch-DNNR	778.15/429.76	R2 = 0.64, MSE = 0.19, MAE = 0.35, MNB = 5.72 1	[94]
							R2 = 0.62, MSE = 0.22, MAE = 0.35, MNB = 7.87, RPD = 1.62 2
NH4+	Shahu Port (CN)	UAV	NA	ML	XGBoost	400–900 nm	R2 = 0.99, RMSE = 0.04, MAE = 0.03 1	[95]
							R2 = 0.95, RMSE = 0.09, MAE = 0.08 2
	Xunsi River (CN)						R2 = 0.91, RMSE = 0.08, MAE = 0.06 1
							R2 = 0.83, RMSE = 0.09, MAE = 0.08 2
NH4+	Poyang Lake (CN)	Landsat 8 (OLI), Sentinel-2 (MSI), Zhuhai-1 OHS-2A (OHS)	NA	ML	LOOCV-GB	Red, NIR	R2 = 0.63, MRE(%) = 14.09%	[96]
	45 Lakes (EE)	Sentinel-2 (MSI)	102	ML	GA-XGBoost	B2 − (B3 + B4)/2, B3 − (B6 + B1)/2, (B6 − B8A) × B5, B2/B4 − B2/B6, B2/B6 − B2/B4, B4/B8A − B4/B1	R2 = 0.99, MAPE(%) = 3.39%, RMSE = 0.00 1	[97]
							R2 = 0.79, MAPE(%) = 75.50%, RMSE = 0.02 2
							R2 = 0.68, MAPE(%) = 161.00%, RMSE = 0.19 3
	Haihe River (CN)	Ground- based	111	NNs	BPNN	400–900 nm	R2 = 0.96, RMSE = 0.25 1	[98]
							R2 = 0.90, RMSE = 0.35 2
AN	Nandu River (CN)	Landsat 8 (OLI)	67	NNs	ANN	B1, B2, B3, B4, B5, B6, B7	R2 = 0.99, RMSE = 0.01, MAPE(%) = 6.09, p < 0.01 1	[99]
							R2 = 0.44, RMSE = 0.19, MAPE(%) = 318.07, p < 0.01 2
NO3−	Zarivar International Wetland (IN)	Landsat 8 (OLI)	NA	NNs	ANN	B4	R2 = 0.28, RMSE = 0.10, MAE = 0.08 1	[100]
							R2 = 0.28, RMSE = 0.70, MAE = 0.04 2
	Nasser Lake (EG)	Sentinel-2 (MSI)	NA	NNs	BWOA-ANN	B1, B2, B3, chl a, TDS—August 2016	NA	[101]
						NA—April 2016
	Bin El Ouidane Reservoir (MA)					B1, B3, B11, chl a—May 2017
	Haihe River (CN)	Ground- Based	111	NNs	BPNN	400–900 nm	R2 = 0.93, RMSE = 0.911	[98]
							R2 = 0.77, RMSE = 2.16 2
	Quinwu, Longjing Nanshan Reservoir (CN)	Sentinel-2 (MSI)	NA	NNs	MDN	B2, B3, B4, B5, B6, B7, B8	NA	[93]
		Landsat 7 (ETM+)				B1, B2, B3, B4
		DJI M300 UAV				B1, B2, B3, B4, B5, B6, B7, B8, B9, B10	R2 = 0.96, RMSE = 0.06, MAE = 0.02 1
							R2 = 0.96, RMSE = 0.06, MAE = 0.02 2
PO43−	45 Lakes (EE)	Sentinel-2 (MSI)	99	ML	GA-XGBoost	B2 × B6/B1, B3 × B6/B2, B7 × B3/B2, B2/(B7 + B6), B2 − (B6 + B4)/2, B5 − (B2 + B3)/2, (B2 − B7) × B4, (B2 − B6)/(B2 − B6), (B2/B6) × (B2/B6)	R2 = 0.99, MAPE(%) = 7.24%, RMSE = 0.00 1	[97]
							R2 = 0.87, MAPE(%) = 43.9%, RMSE = 0.00 2
						B2 × B6/B1, B3 × B6/B2, B7 × B3/B2, B2/(B7 + B6), B2 − (B6 + B4)/2, B5 − (B2 + B3)/2, (B2 − B7) × B4, (B2 − B6)/(B2 − B6), (B2/B6) × (B2/B6)	R2 = 0.45, MAPE(%) = 43.80%, RMSE = 0.00 3
	Zarivar International Wetland (IN)	Landsat 8 (OLI)	NA	NNs	ANN	B3	R2 = 0.49, RMSE = 1.3, MAE = 0.09 1	[100]
							R2 = 0.35, RMSE = 1.28, MAE = 0.85 2
	Nasser Lake (EG)	Sentinel-2 (MSI)	NA	NNs	BWOA-ANN	B1, B3, B4, B5, B8A, chl a—August 2016	NA	[101]
						NA—April 2016
	Bin El Ouidane Reservoir (MA)					B1, B3, B8A, B11, chl a—May 2017
SiO2	Ganga River Basin (IN)	Landsat 8 (OLI)	NA	ML	XGBoost	B1, B2, B3, B4	R = 0.97, R2 = 0.94, Adj. R2 = 0.94	[102]
	45 Lakes (EE)	Sentinel-2 (MSI)	100	ML	GA-XGBoost	B3 × B8A/B4, B4 × B1/B7, B4/B2 − B4/B1, B1/(B7 + B2), (B3 + B5)/B1, (B7 + B2)/B1, B1-(B7 + B6)/2, (B1 − B3) × B5, (B8A − B7) × B6	R2 = 0.99, MAPE(%) = 0.89%, RMSE = 0.03 1	[97]
							R2 = 0.69, MAPE(%) = 168.00%, RMSE = 3.26 2
							R2 = 0.58, MAPE(%) = 123.00%, RMSE = 5.20 3
DO	Weihe River (CN)	SPOT 5	NA	ML	SS-SVR	B1,B2, B3, SWIR	R2 = 1.00, MSE = 0.00	[85]
	Ganga River Basin (IN)	Landsat 8 (OLI)	NA	ML	XGBoost	B1, B2, B3, B4	R = 0.90, R2 = 0.80, Adj. R2 = 0.80	[102]
	Yuhe River (CN)	Groundbased	NA	ML	MLP	Red, Green, NIR	R2 = 1.00, RMSE = 0.23, MAE = 0.16, MAPE(%) = 1.23, Acc.(%) = 100.00% 1	[92]
	Yuhe River (CN)	Ground- based	NA	ML	MLP	Red, Green, NIR	R2 = 0.91, RMSE = 1.16, MAE = 0.99, MAPE(%) = 10.59, Acc.(%) = 94.74% 2	[92]
	Yuandang Lake (CN)	DJI P4 UAV	60	ML	RF	$\mathrm{B}2+\mathrm{B}5$	R2 = 0.96, RMSE = 0.22 1	[91]
							R2 = 0.67, RMSE = 0.62, MAE = 0.41 2
	Q Reservoir (CN)	Sentinel-2 (MSI)	NA	ML	XGBoost	B2, B3, B4, B5, B6, B7, B8, B8A	R2 = 0.90, RMSE = 0.14, MAPE = 0.71, Bias = 0.07	[87]
	Huixian karst Wetland (CN)	Sentinel-2 (MSI)	263	NNs	TR	$\left(\mathrm{B}8+\mathrm{B}2\right)$	R2 = 0.45, RMSE = 0.81	[103]
		Ziyuan-3 (ZY3)				$\left(\mathrm{B}4+\mathrm{B}3\right)$	R2 = 0.59, RMSE = 0.68
		Zhuhai-1 Orbita OHS-01				$\left(\mathrm{B}8+\mathrm{B}11\right)$	R2 = 0.64, RMSE = 0.62
		UAV				$\left(\mathrm{B}5+\mathrm{B}4\right)$	R2 = 0.54, RMSE = 0.72
	Hulum Lake (CN)	Landsat 8 (OLI)	221	ML	RF	(B6 − B7)/(B6 + B7), B6/B7, B6 − B7, B4 − B5, B3 − B4, (B3 − B4)/(B3 + B4), B5 − B7, B1 − B7	R2 = 0.84, MAE = 0.68, RMSE = 0.89	[88]
	45 Lakes (EE)	Sentinel-2 (MSI)	84	ML	GA-XGBoost	B5 × B2/B3, (B4 + B8A) × B3, B5-(B4 + B8A)/2, (B1 − B8A) × B6	R2 = 0.99, MAPE(%) = 1.98%, RMSE = 0.21 1	[97]
DO						B5 × B2/B3, (B4 + B8A) × B3, B5-(B4 + B8A)/2, (B1 − B8A) × B6	R2 = 0.62, MAPE(%) = 15.20%, RMSE = 1.31 2
							R2 = 0.62, MAPE(%) = 46.10%, RMSE = 4.543
	Laspias, Lissos Rivers (GR)	PlanetScope(SuperDove)	NA	ML	SVR	B1, B2, B3, B4, B5, B6, B7, B8	R2 = 0.89, RMSE = 0.71, MAE = 0.53	[104]
							R2 = 0.82, RMSE = 1.41, MAE = 1.07
							R2 = 0.80, RMSE = 0.50, MAE = 0.31
							R2 = 0.65, RMSE = 0.77, MAE = 0.82
							R2 = 0.55, RMSE = 1.82
							R2 = 0.81, RMSE = 1.18, MAE = 0.72
							R2 = 0.69, RMSE = 2.27, MAE = 0.78
	Jingsi Lake, Najing Tonqwei Aquaculture Base (CN)	DJI M300 UAV	50	NNs	TL-Net	NA	R2 = 0.99, MSE = 0.01, RMSE, 0.11, MAE = 0.07, MAPE = 10.94	[105]
	Saint John River (CA)	Landsat 8 (OLI)	38	NNs	BPNN	B1, B2, B3, B4, B5, B6, B7	R2 = 0.99, RMSE = 0.07, p < 0.005 1	[106]
DO							R2 = 0.94, RMSE = 0.19, p < 0.005 2
							R2 = 0.93, RMSE = 0.46, p < 0.005 3
	Mississippi River, Decatur, Carlyle Lake (USA)	Landsat 8 (OLI), Sentinel-2 (MSI), GREON	97	NNs	pDNN	Coastal Blue, Blue, Green, Red, NIR, SWIR	R2 = 0.91, RMSE = 2.06, MAPE = 9.01 1	[107]
							R2 = 0.89, RMSE, 1.81, MAPE = 9.08 2
	Inland water (CN)	DJI M600 pro UAV, WSN	NA	NNs	DNS-DNNs	R850, R632/R550, R590 + R850, (R850 − R632)/R590, (R632 − R590)/R850, (R632 + R590) × R850	R2 = 0.87, RMSE = 0.19 1	[108]
							R2 = 0.80, RMSE = 0.19 2
	Nasser Lake (EG)	Sentinel-2 (MSI)	NA	NNs	BWOA-ANN	B3, B5, B6, chl a—August and April 2016	NA	[101]
	Bin El Ouidane Reservoir (MA)					B5, B8A, B11, chl a—May 2017
	Rawal watershed (stream network) (PK)	Landsat 8 (OLI)	NA	NNs	Bi-LSTM	$\mathrm{B}2/\mathrm{B}4$	R2 = 0.20, MAE = 0.15, MAPE = 0.11	[109]
	Quinwu, Longjing Nanshan Reservoirs (CN)	Sentinel-2 (MSI)	NA	NNs	MDN	B2, B3, B4, B5, B6, B7, B8	NA	[93]
		Landsat 7 (ETM+)				B1, B2, B3, B4
		DJI M300 UAV				B1, B2, B3, B4, B5, B6, B7, B8, B9, B10	R2 = 0.95, RMSE = 0.19, MAE = 0.12 1
							R2 = 0.94, RMSE = 0.24, MAE = 0.12 2

¹ Training set, ² testing set, ³ validation set.

and

Table 4. Remote sensing-based retrieval of nutrients and DO in coastal waterbodies using ML, shallow and deep NN models.

WQP	Study Area	Sensor	N	Model	Algorithm	Bands/Band Indices	Accuracy	Reference
NO3−	Coastal Regions of the East China Sea (CN)	GOCI	193	NNs	BPNN	B1,B2, B3, B4, B5, B6	R2 = 0.98, RMSE = 6.14, MRE(%) = 13.50% 1	[110]
							R2 = 0.98, RMSE = 7.68, MRE(%) = 17.70% 2
							R2 = 0.99, RMSE = 6.13, MRE(%) = 11.20% 3
							R2 = 0.98, RMSE = 6.38, MRE(%) = 13.80% 4
PO43−	Yueqing Bay (CN)	Landsat 8 (OLI)	73	ML	SVR	B4, B5	R2 = 0.86, p < 0.001, Bias: [−0.49, 0.24], MAE(%) = 0.23%, RMSE = 0.00 1	[37]
							R2 = 0.76, p < 0.001, Bias: [−0.36, 0.60], MAE(%) = 0.45%, RMSE = 0.00 2
	Dayu Bay (CN)	Sentinel-2 (MSI)	NA	ML	GPR	B3, B4, B5, B6, B8	R2 = 0.97, RMSE = 3.26 1	[111]
		Sentinel-3 (OLCI)				B6, B8, B11, B12, B17	R2 = 0.60, RMSE = 1.32 2
	Coastal Regions of the East China Sea (CN)	GOCI	193	NNs	BPNN	B1, B2, B3, B4, B5, B6	R2 = 0.86, RMSE = 0.20, MRE(%) = 14.60% 1	[110]
							R2 = 0.75, RMSE = 0.25, MRE(%) = 16.70% 2
							R2 = 0.83, RMSE = 0.22, MRE(%) = 13.3% 3
							R2 = 0.84, RMSE = 0.21, MRE(%) = 14.70% 4
DIN	Yuequing Bay (CN)	Landsat 8 (OLI)	80	ML	SVR	B4, B5, B6, B7	R2 = 0.84, p < 0.001, Bias: [−0.24, 0.32], MAE(%) = 3.48%, RMSE = 0.06 1	[37]
							R2 = 0.81, p < 0.001, Bias: [−0.40, 0.30], MAE(%) = 4.72%, RMSE = 0.06 2
	Dayu Bay (CN)	Sentinel-2 (MSI)	NA	ML	SVR	B3, B4, B5, B6, B8	R2 = 0.67, RMSE = 1.44 1	[111]
DIN		Sentinel-3 (OLCI)				B6, B8, B11, B12, B17	R2 = 0.69, RMSE = 1.33 2
	Coastal Waters, Northern South China Sea (CN)	MODIS	4038	ML	XGBoost	OCR, PCP, STI	R2 = 0.88, MRE = 24.39, RMSE = 0.12	[112]
	Zhejiang Coastal Sea (CN)	MODIS	NA	NNs	ST-DBN	B1, B2, B3, B4, SWIR	R2 = 0.83, RMSE = 0.21, MAE = 0.14 5	[113]
							R2 = 0.84, RMSE = 0.21, MAE = 0.14 6
DIP	Zhejiang Coastal Sea (CN)	MODIS	NA	NNs	ST-DBN	B1, B2, B3, B4, SWIR	R2 = 0.64, RMSE = 0.01, MAE = 0.01 5	[113]
							R2 = 0.65, RMSE = 0.01, MAE = 0.01 6
DO	Lesvos island (GR)	CMEMS	NA	ML	SVR	NA	R2 = 0.32, MAE = 0.11, RMSE = 0.13	[114]
	Shenzhen Bay (CN)	Sentinel-2 (MSI)	64	ML	XGBoost	B2, B3, B4, Red edge, B8, B8A	Error(%) = 0.02%, Bias(%) = 0.00%, Slope = 0.89, RMSLE = 0.07	[115]

¹ Training set, ² testing set, ³ validation set, ⁴ all set, ⁵ sample-based, ⁶ site-based.

.

5.1. Types of Water Bodies

The findings of the literature review suggest that the total number of studies on the retrieval of nutrients and DO pertain to Case II waters, with the majority focusing on highly nutrient-enriched water bodies. Freshwater bodies account for 77% of the total studies, with the majority of them (25% and 24%) referring to lakes and rivers, respectively. Reservoirs, wetlands and stream/stream networks account for 14%, 4% and 3%, respectively. Marine water bodies, which specifically refer to coastal regions, represent 20%, while brackish waters account for only 3% of all studies (Figure 3).

5.2. Remote Sensing Systems for the Retrieval of Nutrients and DO

Different satellite sensors have been used to retrieve nutrients and DO. Since the early 2000s, spectral data from Landsat 4 and Landsat 5 (TM) have been utilized, and over time, additional satellites have been incorporated into research, including IKONOS, Landsat 7 Enhanced Thematic Mapper Plus (ETM+), Moderate Resolution Imaging Spectroradiometer (MODIS) and Satellite pour l’Observation de la Terre 5 (SPOT 5). In recent years, Indian Remote Sensing Linear Imaging Self-Scanning Sensor-III (IRS LISS III), Worldview-2, Landsat 8 Operational Land Imager (OLI), Huang Jing-1 (HJ-1), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), Sentinel-2 Multi-Spectral Instrument (MSI), Landsat 9 Operational Land Imager 2 (OLI-2), Geostationary Ocean Color Imager (GOCI), Ziyuan-3 (ZY-3), Zhuhai-1 OHS-2A (OHS), PlanetScope (SuperDove) and Sentinel-3 Ocean and Land Colour Instrument (OLCI) imagery have also been employed (

Table A1. Overview of spaceborne sensors applied to nutrient and DO estimation.

Multi-spectral	Satellite Sensor	Launch Date	Temporal Resolution (Days)	Spatial Resolution (m)	Spectral Resolution (nm) Band/Wavelength			Reference
	Landsat 4, 5 (TM)	1982, 1984	16	30	B1	450–520	Blue	[118,119]
					B2	520–600	Green
					B3	630–690	Red
					B4	760–900	NIR
					B5	1550–1750	SWIR 1
					B7	2080–2350	SWIR 2
				120	B6	10,410–12,500	TIR
	Landsat 7 (ETM+)	1999	16	15	B8	520–900	Pan	[120]
				30	B1	450–520	Blue
					B2	520–600	Green
					B3	630–690	Red
					B4	770–900	NIR
					B5	1550–1750	SWIR 1
					B7	2090–2350	SWIR 2
				60	B6	10,400–12,500	TIR
	IKONOS	1999	~3	0.82	NA	526–929	Pan	[121]
				3.2	B1	445–516	Blue
					B2	506–595	Green
					B3	632–698	Red
					B4	757–853	NIR
	MODIS	1999	1–2	250	B1	620–670	Red	[122]
					B2	841–876	NIR
				500	B3	459–479	Blue
					B4	545–565	Green
					B5	1230–1250	SWIR
					B6	1628–1652
					B7	2105–2155
				1000	B8	405–420	Violet
					B9	438–448	Blue
					B10	483–493	Blue-Green
					B11	526–536	Green
					B12	546–556
					B13	662–672	Red
					B14	673–683
					B15	743–753	NIR
					B16	862–877
					B17	890–920
					B18	931–941
					B19	915–965
					B26	1360–1390	SWIR
					B20-B36	3660–14,385	TIR
	SPOT 5	2002	2–3	2.5–5	NA	480–710	Pan	[123]
				10	B1	500–590	Green
					B2	610–680	Red
					B3	78–890	NIR
				20	NA	1580–1750	SWIR
	IRS LISS III	2003	24	23.5	B1	520–590	Green	[124]
					B2	620–680	Red
					B3	770–860	NIR
					B4	1550–1700	SWIR
	ASTER	2003	16	15	B1	520–600	VIS	[125]
					B2	630–690
					B3n	760–860	NIR
					B3b	760–860
				30	B4	1600–1700	SWIR
					B5	2145–2185
					B6	2185–2225
					B7	2235–2285
					B8	2295–2365
					B9	2360–2430
				90	B10	8125–8475	TIR
					B11	8475–8825
					B12	8925–9275
B13					10,250–10,950
					B14	10,950–11,650	TIR
	HJ-1	2008	2–3	30	B1	430–520	Blue	[126]
					B2	520–600	Green
					B3	630–690	Red
					B4	760–900	NIR
	WorldView-2	2009	1.1	0.46	NA	450–800	Pan	[127]
				1.8	B1	400–450	Coastal Blue
					B2	450–510	Blue
					B3	510–580	Green
					B4	585–625	Yellow
					B5	630–690	Red
					B6	705–745	Red Edge
					Β7	770–895	NIR 1
					Β8	860–1040	NIR 2
	GOCI	2010	1	500	B1	412	Blue	[128]
					B2	443
					B3	490	Blue-Green
					B4	555	Green
					B5	660	Red
					B6	680
					B7	745	NIR
					B8	865
	ZY-3	2012	5	~2.1	NA	500–800	Pan	[129]
				~3.5–3.6	B1	450–520	Blue
					B2	520–590	Green
					B3	630–690	Red
					B4	770–890	NIR
	Landsat 8 (OLI), Landsat 9 (OLI-2)	2013, 2021	16	15	B8	500–680	Pan	[130,131]
				30	B1	433–453	Coastal/Aerosol
					B2	450–515	Blue
					B3	525–600	Green
					B4	630–680	Red
					B5	845–885	NIR
					B6	1560–1660	SWIR 1
					B7	2100–2300	SWIR 2
					B9	1360–1390	Cirrus
				100	B10	10,600–11,200	TIR 1
					B11	11,500–12,500	TIR 2
	Sentinel-2 (MSI)	2015, 2017	5	10	B2	490	Blue	[132]
					B3	560	Green-peak
					B4	665	Red
					B8	842	NIR
				20	B5	705	Red Edge
					B6	740
					B7	783
					B8a	865	Narrow NIR
					B11	1610	SWIR 1
					B12	2190	SWIR 2
				60	B1	443	Coastal/Aerosol
					B9	940	Water vapor
					B10	1375	Cirrus
	Sentinel-3 (OLCI)	2016, 2018	1–2	300	B1	400	Coastal Blue	[133]
					B2	412.5
					B3	442.5	Blue
					B4	442	Blue-Green
					B5	510	Greenish Cyan
					B6	560	Green
					B7	620	Yellow-Orange
					B8	665	Red
					B9	673.75
					B10	681.25
					B11	708.75	Red Edge
					B12	753.75	NIR
					B13	761.25
					B14	764.38
					B15	767.5
					B16	778.75
					B17	865
					B18	885
	Sentinel-3 (OLCI)	2016, 2018	1–2	300	B19	900	NIR	[133]
					B20	940
					B21	1020	SWIR
	PlanetScope (SuperDove)	2018	1	3	B1	431–452	Coastal Blue	[134]
					B2	465–515	Blue
					B3	513–549	Green
					B4	547–583
					B5	600–620	Yellow
					B6	650–680	Red
					B7	697–713	Red Edge
					B8	845–885	NIR
Hyper-spectral	Zhuhai-1 OHS-2A (OHS)	2018	6	10	B1	443	Violet-Blue	[135]
					B2	466	Blue
					B3	490	Blue-Green
					B4	500	Green
					B5	510	Yellow-Green
					B6	531	Green-Yellow
					B7	550	Yellow
					B8	560	Yellow-Orange
					B9	580	Orange
					B10	596	Orange-Red
					B11	620	Red
	Zhuhai-1 OHS-2A (OHS)	2018	6	10	B12	640	Deep Red	[135]
					B13	665
					B14	670
					B15	686	Red-NIR
					B16	700
					B17	709	NIR
					B18	730
					B19	746
					B20	760
					B21	776
					B22	780
					B23	806
					B24	820
					B25	833
					B26	850
					B27	865
					B28	880
					B29	896
					B30	910
					B31	926
					B32	940

).

The choice of an appropriate sensor depends on the spectral, spatial, temporal and radiometric resolution of the remotely sensed images. The cost of image acquisition is also crucial [16,33]. In small water bodies, high-resolution imagery is required [51,104]. High-resolution satellites (e.g., IKONOS, Worldview-2, PlanetScope (SuperDove), Sentinel-2 MSI) provide detailed spatial information in optically complex water bodies, where highly variable features are present. Moreover, they have the potential to minimize the mixed-pixel effect. In small water bodies, mixed could be observed when low-resolution imagery is utilized. On the other hand, high-resolution satellites are not suitable for large water bodies due to their limited spatial coverage and lower temporal resolution. For regional and global scale studies, moderate- and low-resolution satellites (e.g., MODIS and GOCI) are preferred [80,110,112,113].

Several studies compared the accuracy of different satellite sensors in retrieving nutrients and DO. Satapathy et al. [78] employed data from Landsat 7 (ETM+) and IKONOS to quantitatively retrieve PO₄³⁻ in the West Coast of Mumbai (IN). The results indicated high prediction accuracy using either set of satellite data.

Khattab and Merkel [57] analyzed remote sensing data obtained from Landsat 5 (TM) and Landsat 7 (ETM+) imagery to retrieve NO₂⁻, NO₃⁻ and PO₄³⁻ in Mosul Dam Lake (IQ). The models based on Landsat 7 (ETM+) spectral data showed more precise results compared to Landsat 5 (TM). Pizani et al. [75] compared the performance of Sentinel-2 (MSI) and Landsat 8 (OLI) for the retrieval of DO in Três Marias Reservoir (BR). The utilized reflectance values were based on averaging kernels of 30 m and 90 m. Sentinel-2 (MSI) achieved higher predictive accuracy for both kernel sizes compared to Landsat 8 (OLI). Fu et al. [96] used Sentinel-2 (MSI) and Zhuhai-1 OHS-2A (OHS) reflectance data to quantitatively retrieve NH₄⁺ in Poyang Lake (CN). Hyperspectral images from Zhuhai-1 OHS-2A (OHS) achieved better predictive results compared to the multispectral images from Sentinel-2 (MSI).

Researchers have also utilized data from multiple satellite sensors to achieve higher predictive accuracy. Peterson et al. [107] used a harmonized virtual constellation of Landsat 8 (OLI) and Sentinel-2 (MSI) to create a high temporal frequency dataset at a relatively fine spatial scale. Elhag et al. [60] used a spectral correspondence between Sentinel-2 (MSI) and Sentinel-3 (OLCI). The correlation between all bands was determined and the bands with similar central wavelengths were selected to construct the retrieval models (Table 1, Table 2, Table 3 and Table 4).

Recently, with the rapid development of Unmanned Aerial Vehicle (UAV) technology, drones have also been used in the retrieval of nutrients and DO using either multispectral [77,90,91,93,108] or hyperspectral imaging [94,95] (

Table A2. Overview of UAV sensors applied to nutrient and DO estimation.

	Airborne UAV Platform	Spatial Resolution (m)	Spectral Resolution (nm) Band/Wavelength			Reference
Multi-spectral	DJI M600 pro	0.155	NA	395–1000	NA	[108]
	Red edge-MX	0.1	B1	475	Blue	[90]
			B2	560	Green
			B3	670	Red
			B4	720	Red Edge
			B5	840	NIR
	DJI P4, DJI P4M, DJI Elf 4	NA, 0.06, NA	B1	450	Blue	[77,91,103]
			B2	560	Green
			B3	650	Red
			B4	730	Red Edge
			B5	840	NIR
	DJI M300 RTK	0.15	B1	444	Coastal Blue	[93]
			B2	475	Blue
			B3	531	Green
			B4	560
			B5	650	Red
			B6	668
			B7	705	Red Edge
			B8	717
			B9	740
			B10	842	NIR
Hyper-spectral	DJ M600 pro	0.185	NA	400–900	NA	[95]
	NA	2	NA	400–1000	NA	[94]
	DJI M300	0.12	NA		NA	[105]

). Tian et al. [93] used the reflectance values of DJI M300 RTK UAV multispectral platform, Sentinel-2 (MSI) and Landsat 7 (ETM+) to estimate NO₃⁻, NH₃ and DO in the Quinwu, Longjing and Nanshan Reservoirs (CN). The accuracy of the model derived from the airborne UAV dataset was relatively higher compared to spaceborne satellite sensors. Unlike satellites, drones can capture the continuous spectral signatures of water bodies, enabled by their flexible and on-demand deployment capabilities [15]. In addition to satellite and UAV sensors, ground-based spectrometers [29,92,98] have been used as an alternative source of remotely sensed data (Table 1, Table 2, Table 3 and Table 4). UAV and ground-based sensors provide data with higher spectral and spatial resolution, while are much less affected by the atmosphere compared to satellites. However, they have the limitations of high cost and limited capacity to observe large water bodies [15,33].

5.3. Remote Sensing Methods for the Retrieval of Nutrients and DO

As mentioned in Chapter 1.5, there are four different approaches by which WQPs can be extracted using remotely sensed data. However, in the case of nutrients and DO, researchers have focused on two types of methods, the empirical and the AI method. Based on the existing literature, the empirical approach is the most commonly used, representing 56% (58 studies) of the total. AI-based methods follow (48%), comprising 32 ML and 21 NN studies, respectively (Figure 4). Analytical and semi-empirical approaches appear not to be used in establishing inversion models for nutrients and DO. Table 1, Table 2, Table 3 and Table 4 include a detailed list of the empirical and AI methods used for the retrieval of nutrients and DO, as well as their prediction accuracy based on various statistical indices.

5.3.1. Empirical Models

Empirical models are simple and easy to operate. However, they require a large number of in situ measurements to assure a higher accuracy. In addition, they have poor generalization abilities in the spatial and temporal scales [45,116] and usually are not suitable for applications in Case II waters due to the various constituents that affect the optical properties of the water [44]. Empirical approaches employ regression analysis using various polynomial and non-polynomial equations. Among these, multiple linear regression (MLR), stepwise MLR and simple linear regression (SLR) are the most widely applied techniques for nutrient and DO estimation (Figure 5; Table 1 and Table 2).

Based on Sentinel-2 (MSI) data, Dong et al. [51] applied SLR to establish inversion models for NH₃ estimation in Danjiangkou Reservoir (CN). The quadratic model had the highest prediction accuracy (R² = 0.85, MAE = 0.01, RMSE = 0.01) compared to the other constructed empirical models. The band ratio between the green peak (560 nm) and the blue (490 nm) band was formed to develop the retrieval model. A quadratic model was, also, developed by Abdelmalik [65] to estimate PO₄³⁻ in Quaroum Lake (EG), using the SWIR (2145–2185 nm) band of ASTER. By using Landsat’s 8 (OLI) data, Markogianni et al. [54,55] applied several empirical models to estimate NH₄⁺ concentration in Trichonis Lake (GR). The cubic model showed a higher prediction accuracy compared to the other models, with weak (R² = 0.47, SEE = 0.00) prediction results. The band combination of the coastal blue (433–453 nm), green (525–600 nm) and red (630–680 nm) bands was used to establish the inversion model. For NO₃⁻, NH₄⁺, SiO₂ and PO₄³⁻ estimation in the Xiangxi River (CN), Liu et al. [53] built several empirical models utilizing HJ-1 satellite imagery. The cubic and the quadratic model provided moderate predictive accuracy for NO₃⁻ (R² = 0.75) and SiO₂ (R² = 0.56), respectively, utilizing a band combination of the blue (430–520 nm), green (520–600 nm) and NIR (760–900 nm) bands. The retrieval models of NH₄⁺ and PO₄³⁻ exhibited weak predictive accuracy (R² < 0.5), using a band combination of the green, red (630–690 nm) and NIR bands and a band combination of the blue and NIR bands, respectively. Qui et al. [69] established an exponential model for the inversion of DO in the Huangpu River (CN), with moderate predictive accuracy (R² = 0.68), using the band ratio between the NIR (760–900 nm) and the green (520–600 nm) bands of Landsat 5 (TM) (Table 1 and Table 2).

5.3.2. ML Models

The most commonly applied regression-based ML algorithms in studies related to nutrient and DO estimation are the following:

• Support Vector Regression (SVR) [37,85,86,87,89,92,93,94,99,102,104,105,106,107,111,114,115];
• Extreme Gradient Boosting (XGBoost) [87,90,93,95,96,97,102,103,105,112,115];
• Categorical Boosting (CatBoost) [96,105];
• Gradient Boosting (GB) [91,92,96,99,103];
• Random Forest (RF) [87,88,89,90,91,92,96,99,102,103,112];
• Adaptive Boosting (AdaBoost) [90,92,96];
• Regression Tree (RT) [99];
• Elastic Net Regression (ENR) [92];
• k-Nearest Neighbor Regression (k-NNR) [92];
• Gaussian Process Regression (GPR) [92,111];
• Extra Trees Regression (ETR) [92];
• Extreme Learning Machine Regression (ELM) [107,108].

Among the ML algorithms reviewed, XGBoost [87,95,102,112,115] and SVR [11,37,104,114] exhibit the highest predictive accuracy (Figure 5; Table 3 and Table 4).

5.3.3. NN Models

The most widely acknowledged NN architectures employed in the retrieval of nutrients and DO are the following:

• Backpropagation Neural Network (BPNN) [89,91,98,106,110];
• Artificial Neural Network (ANN) [87,99,100,101,108];
• Multilayer Perceptron (MLP) [92,102,109,115];
• Pixel based-Deep Neural Network Regression (pixel-DNNR) [94];
• Patch based-Deep Neural Network Regression (patch-DNNR) [94];
• Binary Whale Optimization Algorithm-Artificial Neural Network (BWOA-ANN) [101];
• Transformer (TR) [103,105];
• Deep learning model that integrates Transformer and LSTM Networks (TL-Net) [105];
• Convolutional Neural Network (CNN) [91,105,109];
• Fully Connected Neural Network (FCNN) [109];
• Recurrent Neural Network (RNN) [109];
• Long Short-Term Memory (LSTM) [105];
• Vanilla-Long Short-Term Memory (V-LSTM) [109];
• Stacked-Long Short-Term Memory (S-LSTM) [109];
• Bidirectional-Long Short-Term Memory (Bi-LSTM) [109];
• Convolutional-Long Short-Term Memory (Conv-LSTM) [109];
• Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) [109];
• Mixture Density Network (MDN) [93,103,115];
• Progressively decreasing Deep Neural Network (pDNN) [107];
• Dynamic Network Surgery-Deep Neural Network (DNS-DNNS) [108];
• Deep Neural Network (DNN) [90,93];
• Spatiotemporal-Deep Belief Network (ST-DBN) [113];
• Message Passing Neural Network (MPNN) [113];
• Generalized Regression Neural Network (GRNN) [113].

The BPNN [98,106,110] and ANN [99,100] are the most commonly adopted models demonstrating high prediction accuracy (Figure 5; Table 3 and Table 4).

5.3.4. Comparative Analysis of Predictive Accuracy: Empirical, ML and NN Models

The accuracy of different retrieval models, including empirical, ML and NN approaches, has been compared in several studies.

Empirical vs. ML Models

Similarly to empirical models, ML models are only applicable within the range and settings of the training data. However, this approach uses iterative learning to reduce the overall error and maximize the model fit. Furthermore, the ML is able to operate in multidimensional space and capture complex, non-linear relationships [45].

Based on SPOT 5 imagery, Wang et al. [85] developed a MLR and a semi-supervised SVR (SS-SVR) model to retrieve NH₃ and DO in Weihe River (CN). SS-SVR showed higher prediction accuracy (R² = 0.98, MSE = 0.05 for NH₃ and R² = 1.00, MSE = 0.00 for DO) compared to MLR (R² = 0.48, MSE = 4.57 for NH₃ and R² = 0.81, MSE = 0.92 for DO). Similarly, Wang et al. [86] developed a MLR model and a genetic algorithm (GA) SVR model (GA-SVR) to quantitatively retrieve NH₃. The SVR model indicated better prediction accuracy (R² = 0.98, MSE = 0.77, MAD = 0.43) compared to the MLR (R² = 0.76, MSE = 4.46, MAD = 3.03). In both studies, the single bands from the green (500–590 nm) to SWIR (1580–1750 nm) were used as independent variables to construct the inversion models. As indicated by the authors, the results from a small number of in situ measurements showed that SVR could provide more accurate predictive results compared to traditional MLR. Wang et al. [95] applied various empirical models and the XGBoost model to estimate NH₄⁺ in Shahu Port and Xunsi River (CN), using UAV hyperspectral imagery. Spectral data in the 400 to 900 nm range of the electromagnetic spectrum were used to construct the retrieval models. XGBoost demonstrated improved prediction accuracy compared to traditional empirical models (Table 3).

Empirical vs. NN Models

Amanollahi et al. [100] used an ANN model and a linear regression (LR) model to retrieve NO₃⁻ and PO₄³⁻ in Zarivar International Wetland (IR) based on Landsat 8 (OLI) image data. The results showed that for both the ANN and LR model R², RMSE and MAE values exhibited low predictive accuracy. On the other hand, Wu et al. [113] established a DBN, MPNN, GRNN and a MLR model in both original and spatio-temporal-integrated forms to estimate DIN and DIP in the Zhenjiang Coastal Sea (CN). MODIS’s single bands from red (620–670 nm) to SWIR 3 (2105–2155 nm), except SWIR 2 (1628–1652 nm), were extracted and used as input features to the retrieval models. The spatio-temporal-integrated deep learning framework achieved significantly higher R² values and reduced estimation errors compared to the original forms. The ST-DBN model showed the highest predictive accuracy for both DIN (R² = 0.83, RMSE = 0.21, MAE = 0.14) and DIP (R² = 0.64, RMSE = 0.01, MAE = 0.01), while the ST-MLR model was the worst (R² = 0.62, RMSE = 0.32, MAE = 0.20 for DIN, R² = 0.42, RMSE = 0.02, MAE = 0.01 for DIP) (Table 3 and Table 4).

ML vs. NN Models

Theoretically, in the case of large and high-dimensional data, NNs outperform conventional empirical and ML approaches [46]. However, the complex architecture of NNs combined with the high computational performance and the huge amount of training data required make their use more complicated compared to ML models [48]. On the other hand, in the case of limiting and low-dimensional data, ML tends to provide better results than those generated by NNs [46].

Based on Landsat 8 (OLI) data, Sharaf El Din et al. [106] established a BPNN model and an SVR model for estimating DO, in Saint John River (CA). The BPNN model showed higher accuracy (R² = 0.94, RMSE = 0.19) compared to the SVR model (R² = 0.89, RMSE = 0.57). Single bands, from the coastal blue (433–453 nm) to SWIR 2 (2100–2300 nm), except for the green (525–600 nm) band, were selected to form the input layer of the inversion model. Li et al. [99] used four different ML algorithms (SVR, RF, RT and GBM) and an ANN model to retrieve AN in Nandu River (China), using Landsat 8 (OLI) imagery. The spectral data from coastal blue to SWIR 2 were used as input parameters for the retrieval models. The results demonstrated that the ANN model had the highest R² values, which were close to 1.00 for the training and 0.44 for the testing dataset. The RMSE values were 0.01 and 0.19, respectively. Other researchers compared the prediction accuracy of the XGBoost algorithm with the SVR, MLP, MDN, RF and ANN models for DO and NH₃ estimation, using the single bands from the blue (490 nm) to NIR narrow (865 nm) of Sentinel-2 (MSI) in Shenzhen Bay and Q Reservoir (CN), respectively. In both case studies, the results indicated that the XGBoost outperformed the other ML and NN models and can provide better prediction accuracy when limited sample data are available [87,115]. Lo et al. [91] used the RF, GB, BPNN and CNN models to estimate DO and NH₃ in Yuandang Lake (CN), using UAV multispectral imagery. In the case of DO, RF provided higher accuracy results both for the training (R² = 0.96, RMSE = 0.22) and the testing (R² = 0.67, RMSE = 0.62, MAE = 0.41) datasets, using the band combination of the green (560 nm) and NIR (840 nm) bands. For NH₃ estimation, GB showed the best prediction accuracy (R² = 0.84, RMSE = 0.06 for training and R² = 0.55, RMSE = 0.12, MAE = 0.09 for testing dataset) compared to the RF and the DL models. The band combination of the red band (650 nm) and red edge (730 nm) was selected as the independent variable to establish the retrieval model (Table 3 and Table 4).

Empirical vs. ML vs. NN Models

Using the spectral reflectance data derived from the harmonized virtual constellation of Landsat 8 (OLI) and Sentinel-2 (MSI), Peterson et al. [107] applied a pDNN model for the estimation of DO, in Mississippi River, Lake Decatur and Lake Carlyle (USA). The developed model was evaluated against MLR, SVR and ELM. The pDNN model outperformed all the other retrieval models (R² = 0.91, RMSE = 2.06, MAPE = 9.01 for training dataset, R² = 0.89, RMSE = 1.81, MAPE = 9.08 for testing dataset), while the MLR model generated the lowest prediction accuracy (R² = 0.93, RMSE = 13.65, MAPE = 8.90 for training dataset, R² = 0.44, RMSE = 3.31, MAPE = 13.98 for testing dataset). Based on Sentinel-2 (MSI), Landsat 7 (ETM+) and UAV multispectral imagery, Tian et al. [93] constructed an unclustered band combination (UC-BC) empirical model, fuzzy C-means band combination (FCM-BC) empirical model, XGBoost, SVR, MDN and DNN to quantitatively retrieve NO₃⁻, NH₃ and DO in the Quinwu, Longjing and Nanshan Reservoirs (CN). The prediction accuracy of FCM-BC improved compared to the BC model; however, it lagged behind the ML and DL models. The MDN model showed the best prediction accuracy. Chen et al. [90] compared the prediction accuracy between GA-XGBoost, RF, GA-RF, AdaBoost, GA-AdaBoost, DNN and several empirical models for NH₃ estimation in the Nanfei River (CN), using UAV multispectral imagery. The performance of GA-XGBoost (R² = 0.69, MAE = 0.14, RMSE = 0.16) was better compared to the empirical, DNN and other ML models. The band combinations of the blue (475 nm), green (560 nm) and red (670 nm) bands and red edge (720 nm) were used as input features of the GA-XGBoost model (Table 3).

5.4. Parameters That Influence the Accuracy of the Retrieval Models

In optically shallow waters (depth less than 30 m) the water-leaving radiance may be influenced by the light reflected off the bottom of the water body. This influence could vary depending on water clarity and bottom composition (e.g., sand, mud, rock, coral reefs, seagrass and macroalgae). The bottoms of Case II water bodies often exhibit strong influences and more complex models may be required to account for the effects of depth and seabed on ocean color [12,34,35]. In order to establish the relationship between NH₄⁺, NO₃⁻, DO and reflectance data from Landsat 7 (ETM+) and Landsat 8 (OLI), Theologou et al. [52] developed three simple linear regression models, each corresponding to a specific depth of a shallow freshwater body. The first model referred to relatively deep parts, the second to very shallow parts and the third to all lake depths. In general, the single linear regression models did not establish well-defined relationships between WQPs and remotely sensed data and only the first model provided accurate predictive results. Moreover, the shallow parts of the water body appeared to follow different regression patterns.

Seasonal variations may influence the spectra detected by a remote sensor [16,34,65,75]. Seasonal in situ measurements and Landsat 8 (OLI) spectral data were used by Khalil et al. [72] to quantitatively retrieve DO in Bardawil Lagoon (EG). Stepwise multiple linear regression models were developed, indicating different prediction accuracy between the four seasons. Similar, Vatitsi et al. [104] employed two different experimental setups to determine the impact of seasonality in the retrieval of DO using PlanetScope (SuperDove) spectral data. One predictive model was developed, considering data from multiple seasons, while two models were developed focusing on spring and summer, respectively. The accuracy between the two experimental setups varied, while the two seasonal models demonstrated different levels of performance. Besides seasonal variations, different concentration levels of nutrients and DO could affect the accuracy of the retrieval models. Indeed, Xu et al. [79] mentioned that it is not feasible to retrieve DIN via remote sensing when concentration is below 70 μg/L (Table 1, Table 2 and Table 3).

6. Discussion and Conclusions

In the present study, a systematic review was conducted to highlight recent advances and the current state of the art of quantitative nutrient and DO concentration retrieval using remote sensing techniques. A narrative literature synthesis was undertaken due to the heterogeneity among studies, and findings were summarized qualitatively. The findings were limited to Case II waters, where both empirical and AI models were applied. Studies were grouped by WQP, study area (inland or coastal waters) and retrieval model (empirical, ML or NNs) to identify patterns in retrieval accuracy based on the reported validation metrics.

Differences in sample size, validation processes and validation metrics limited our ability to make direct comparisons between studies. Some studies used entire datasets for model training without employing independent testing, while others did not clarify whether the reported metrics were derived from independent training or testing data. In addition, validation metrics varied across studies, and in some cases, sample size information was not reported. This lack of consistency constrained the strength of general conclusions.

Patterns in retrieval accuracy were assessed solely by comparing results within the same study area with the same number of samples and using comparable validation processes and metrics. In such cases, ML and NNs outperformed empirical models [85,86,93,94,95,105,107,113]. In general, empirical models require a large number of in situ measurements to achieve high accuracy [45,116], while NNs also demand substantial training datasets [46]. However, the exact sample size needed for either approach is not clearly defined in the literature. ML models (e.g., SVR and XGBoost) often outperform NNs when training data are limited, as they are generally less prone to overfitting under small-sample conditions [46].

Although several studies report high predictive accuracy, current models are site-specific, and their transferability across different regions may be restricted. To improve their transferability, existing models should be validated across diverse regions and time periods. Moreover, they should be applied across a wide range of environmental conditions, considering trophic state (e.g., oligotrophic, mesotrophic, eutrophic and hypereutrophic conditions), as well as seasonal changes in nutrient and DO concentration. Additionally, measurements of the water column profile might be helpful. In optically shallow waters, depth and bottom characteristics could also be incorporated into models due to their influence on the signals detected by remote sensors [117]. In almost all cases, regression models were used; however, if a wide range of training datasets is available, cluster-based models could also be tested in future.

Remote sensing shows strong potential as a complementary tool for monitoring non-optically active WQPs such as nutrients and DO. However, several challenges remain, and therefore further research is required. Future work should emphasize transparent reporting of sample size and validation metrics to improve the reliability and comparability of results.