Sun-induced fluorescence (SIF) from phytoplankton has historically been used as a proxy for chlorophyll-a concentration (chl-a) estimates in water bodies using optical Earth observation data. However, the relationship is often affected by spectral features caused by elastic scattering, and by the shifting incidence of different fluorescence quenching mechanisms. This study found that disentangling Photochemical Quenching (PQ) and Non-Photochemical Quenching (NPQ) cases improves SIF-based chl-a estimates. Further, we defined strategies which can distinguish the two quenching mechanisms. We assembled a unique dataset collected between 2018 and 2022 by an autonomous profiler in Lake Geneva (Western Europe). We used NPQ-influenced chl-a estimates from the fluorometer and NPQ-corrected chl-a estimates to distinguish between PQ and NPQ cases. The correlation between SIF yield and chl-a is weak when considering the entire dataset (R² = 0.37; Median Absolute Percentage Difference (MAPD) = 74%). It increases strongly when comparing PQ cases (R² = 0.72, MAPD = 49%) and NPQ cases (R² = 0.48, MAPD = 68%) separately. Analyzing a subset of in situ measurements acquired around Sentinel-3 overpasses (±3 hours) improved the performance metrics for both PQ cases (R² = 0.82, MAPD = 35%) and NPQ cases (R² = 0.43, MAPD = 61%). However, when applying the same approach to Sentinel-3 OLCI data, we found that the errors in remote sensing reflectance products disable such an adaptation. We conclude that enhanced atmospheric correction in the red-NIR region for oligo-mesotrophic lakes is needed to demonstrate the upscaling of our in situ-based results. This will enhance satellite-based SIF yield retrievals and subsequently, obtain SIF-related phytoplankton physiology products.

Machine learning models have steadily improved in estimating inherent optical properties (IOPs) from remote sensing observations. Yet, their generalization ability when applied to new water bodies, beyond those they were trained on, is not well understood. We present a novel approach for assessing model generalization across various scenarios, including interpolation within in situ observation datasets, extrapolation beyond the training scope, and application to hyperspectral observations from the PRecursore IperSpettrale della Missione Applicativa (PRISMA) satellite involving atmospheric correction. We evaluate five probabilistic neural networks (PNNs), including novel architectures like recurrent neural networks, for their ability to estimate absorption at 443 and 675 nm from hyperspectral reflectance. The median symmetric accuracy (MdSA) worsens from ≥25% in interpolation scenarios to ≥50% in extrapolation scenarios, and reaches ≥80% when applied to PRISMA satellite imagery. Across all scenarios, models produce uncertainty estimates exceeding 40%, often reflecting systematic underconfidence. PNNs show better calibration during extrapolation, suggesting an intrinsic awareness of retrieval constraints. To address this miscalibration, we introduce an uncertainty recalibration method that only withholds 10% of the training dataset, but improves model calibration in 86% of PRISMA evaluations with minimal accuracy trade-offs. Resulting well-calibrated uncertainty estimates enable reliable uncertainty propagation for downstream applications. IOP retrieval uncertainty is predominantly aleatoric (inherent to the observations). Therefore, increasing the number of measurements from the same distribution or selecting a different neural network architecture trained on the same dataset does not enhance model accuracy. Our findings indicate that we have reached a predictability limit in retrieving IOPs using purely data-driven approaches. We therefore advocate embedding physical principles of IOPs into model architectures, creating physics-informed neural networks capable of surpassing current limitations.

Bloom-forming algae present a unique challenge to water managers as they can significantly impair provision of important ecosystem services and cause health risks to humans and animals. Consequently, effective short-term algae forecasts are important as they provide early warnings and enable implementation of mitigation strategies. In this context, machine learning (ML) emerges as a promising forecasting tool. However, the performance of ML models is heavily dependent on the availability of appropriate training data. Consequently, it is essential to determine the volume of data necessary to develop reliable ML forecasts. Understanding this will guide future monitoring strategies, optimize resource allocation, and set realistic expectations for management outcomes. In this study, we used 30 years of fortnightly measurements of 13 different parameters from a lake in the English Lake District (UK) to examine the impact of training data duration on the performance of ML models for forecasting chlorophyll-a two weeks in advance. Once training data availability exceeded four years, a Random Forest model was found to consistently outperform naive benchmarks (mean absolute percentage error 16.4 % lower than the best-performing benchmark). With more than 5 years of training data, model performance generally continued to improve, but with diminishing returns. Furthermore, it was found that equivalent and, in some cases, better performance could be achieved by only using a subset of the most important input features. Additionally, it was found that reducing the sampling frequency had negative impacts on performance, both due to the reduced number of training observations available, and increased forecast horizon. Our findings demonstrate that for lakes ecologically similar to the study site, a consistent and regular sampling programme focused on monitoring a limited number of key parameters can provide sufficient observations for generating short-term algae forecasts after approximately five years of data collection. Importantly, this result provides justification for the initiation of new monitoring programmes for sites where algal blooms are a concern, and suggests that there are likely many pre-existing monitoring datasets which would be suitable for training algae forecast models.

Freshwater algae exhibit complex dynamics, particularly in meso-oligotrophic lakes with sudden and dramatic increases in algal biomass following long periods of low background concentration. While the fundamental prerequisites for algal blooms, namely light and nutrient availability, are well-known, their specific causation involves an intricate chain of conditions. Here we examine a recent massive Uroglena bloom in Lake Geneva (Switzerland/France). We show that a certain sequence of meteorological conditions triggered this specific algal bloom event: heavy rainfall promoting excessive organic matter and nutrients loading, followed by wind-induced coastal upwelling, and a prolonged period of warm, calm weather. The combination of satellite remote sensing, in-situ measurements, ad-hoc biogeochemical analyses, and three-dimensional modeling proved invaluable in unraveling the complex dynamics of algal blooms highlighting the substantial role of littoral-pelagic connectivities in large low-nutrient lakes. These findings underscore the advantages of state-of-the-art multidisciplinary approaches for an improved understanding of dynamic systems as a whole.

Given the use of machine learning-based tools for monitoring the Water Quality Indicators (WQIs) over lakes and coastal waters, understanding the properties of such models, including the uncertainties inherent in their predictions is essential. This has led to the development of two probabilistic NN-algorithms: Mixture Density Network (MDN) and Bayesian Neural Network via Monte Carlo Dropout (BNN-MCD). These NNs are complex, featuring thousands of trainable parameters and modifiable hyper-parameters, and have been independently trained and tested. The model uncertainty metric captures the uncertainty present in each prediction based on the properties of the model—namely, the model architecture and the training data distribution. We conduct an analysis of MDN and BNN-MCD under near-identical conditions of model architecture, training, and test sets, etc., to retrieve the concentration of chlorophyll-a pigments (Chl a), total suspended solids (TSS), and the absorption by colored dissolved organic matter at 440 nm (a_cdom (440)). The spectral resolutions considered correspond to the Hyperspectral Imager for the Coastal Ocean (HICO), PRecursore IperSpettrale della Missione Applicativa (PRISMA), Ocean Colour and Land Imager (OLCI), and MultiSpectral Instrument (MSI). The model performances are tested in terms of both predictive residuals and predictive uncertainty metric quality. We also compared the simultaneous WQI retrievals against a single-parameter retrieval framework (for Chla). Ultimately, the models’ real-world applicability was investigated using a MSI satellite-matchup dataset (N = 3'053) of Chla and TSS. Experiments show that both models exhibit comparable estimation performance. Specifically, the median symmetric accuracy (MdSA) on the test set for the different parameters in both algorithms range from 30% to 60%. The uncertainty estimates, on the other hand, differ strongly. MDN’s uncertainty estimate is ∼50%, encompassing estimation residuals for 75% of test samples, whereas BNN-MCD’s average uncertainty estimate is ∼25%, encompassing the residuals for 50% of samples. Our analysis also revealed that simultaneous estimation results in improvements in both predictive performance and uncertainty metric quality. Interestingly, the trends mentioned above hold across different sensor resolutions, as well as experimental regimes. This disparity calls for additional research to determine whether such trends in model uncertainty are inherent to specific models or can be more broadly generalized across different algorithms and sensor setups.

The removal of organic particulate matter, predominantly phytoplankton, in eutrophic coastal seas and estuaries is considered an ecosystem service performed by large bivalve assemblages. Mussel farming has been proposed as a measure to mitigate eutrophication, as filtration directly reduces the concentration of chlorophyll a (Chl a), a primary ecological indicator. Seston depletion is typically assessed by in situ investigation, which generally lacks spatiotemporal coverage of features relative to greater ecosystem dynamics. To assess the scale and structure of this service, the present study couples multiple measurement approaches, including moored stations, synoptic transect surveys, flow cytometry, a preliminary drone survey technique, and satellite remote sensing within and around a large mussel farm. Significant depletion patterns were observed with all methods, and mixing gradients could be detected hundreds of meters beyond the farm, with repeatable patterns but distinct findings between methods. The intensity of the depletion signal was correlated with mussel biomass loads, ambient conditions, and hydrodynamic regimes, ranging from 5% to 91% relative Chl a depletion and Secchi depth increases of up to 2 m. Changes in particle size distributions were impacted in all downstream areas, as well as phytoplankton diversity. Observed depletion gradients with satellite imagery were consistent with other measurements and can be used to complement in situ field measurements. The findings of this study, will inform carrying capacity assessments, farm configuration, and development of impact assessment programs on seston removal for bivalve aquaculture.

Inland and coastal waters provide key ecosystem services and are closely linked to human well-being. In this study, we propose a semi-analytical method, which can be applied to Sentinel-2 MultiSpectral Instrument (MSI) images to retrieve high spatial-resolution total suspended solids (TSS) concentration in a broad spectrum of aquatic ecosystems ranging from clear to extremely turbid waters. The presented approach has four main steps. First, the remote sensing reflectance (R_rs) at a band lacking in MSI (620 nm) is estimated through an empirical relationship from R_rs at 665 nm. Second, waters are classified into four types (clear, moderately turbid, highly turbid, and extremely turbid). Third, semi-analytical algorithms are used to estimate the particulate backscattering coefficient (b_bp) at a reference band depending on the water types. Last, TSS is estimated from b_bp at the reference band. Validation and comparison of the proposed method with three existing methods are performed using a simulated dataset (N = 1000), an in situ dataset collected from global inland and coastal waters (N = 1265) and satellite matchups (N = 40). Results indicate that the proposed method can improve TSS estimation and provide accurate retrievals of TSS from all three datasets, with a median absolute percentage error (MAPE) of 14.88 %, 31.50 % and 41.69 % respectively. We also present comparisons of TSS mapping between the Sentinel-3 Ocean and Land Colour Instrument (OLCI) and MSI in Lake Kasumigaura, Japan and the Tagus Estuary, Portugal. Results clearly demonstrate the advantages of using MSI for TSS monitoring in small water bodies such as rivers, river mouths and other nearshore waters. MSI can provide more detailed and realistic TSS estimates than OLCI in these water bodies. The proposed TSS estimation method was applied to MSI images to produce TSS time-series in Lake Kasumigaura, which showed good agreements with in situ and OLCI-derived TSS time-series.

The development of algorithms for remote sensing of water quality (RSWQ) requires a large amount of in situ data to account for the bio-geo-optical diversity of inland and coastal waters. The GLObal Reflectance community dataset for Imaging and optical sensing of Aquatic environments (GLORIA) includes 7,572 curated hyperspectral remote sensing reflectance measurements at 1 nm intervals within the 350 to 900 nm wavelength range. In addition, at least one co-located water quality measurement of chlorophyll a, total suspended solids, absorption by dissolved substances, and Secchi depth, is provided. The data were contributed by researchers affiliated with 59 institutions worldwide and come from 450 different water bodies, making GLORIA the de-facto state of knowledge of in situ coastal and inland aquatic optical diversity. Each measurement is documented with comprehensive methodological details, allowing users to evaluate fitness-for-purpose, and providing a reference for practitioners planning similar measurements. We provide open and free access to this dataset with the goal of enabling scientific and technological advancement towards operational regional and global RSWQ monitoring.

Uncertainty is an inherent aspect of aquatic remote sensing, originating from sources such as sensor noise, atmospheric variability, and human error. Although many studies have advanced the understanding of uncertainty, it is still not incorporated routinely into aquatic remote sensing research. Neglecting uncertainty can lead to misinterpretations of results, missed opportunities for innovative research, and a limited understanding of complex aquatic systems. In this article, we demonstrate how working with uncertainty can advance remote sensing through three examples: validation and match-up analysis, targeted improvement of data products, and decision-making based on information acquired through remote sensing. We advocate for a change of perspective: the uncertainty inherent in aquatic remote sensing should be embraced, rather than viewed as a limitation. Focusing on uncertainty not only leads to more accurate and reliable results but also paves the way for innovation through novel insights, product improvements, and more informed decision-making in the management and preservation of aquatic ecosystems.

Satellite remote sensing of chlorophyll-a concentration (chla) in oligotrophic and mesotrophic lakes faces uncertainties from sources such as atmospheric correction, complex inherent optical property compositions, and imperfect algorithmic retrieval. To improve chla estimation in oligo- and mesotrophic lakes, we developed Bayesian probabilistic neural networks (BNNs) for the Sentinel-3 Ocean and Land Cover Instrument (OLCI) and Sentinel-2 MultiSpectral Imager (MSI). The BNNs were built using an in situ dataset of oligo- and mesotrophic water bodies (1755 observations from 178 systems; median chla: 5.11 mg m^-3, standard deviation: 10.76 mg m^-3) and provide a per-pixel uncertainty percentage associated with retrieved chla. Shifts of oligo- and mesotrophic systems into the eutrophic regime, characterised by higher biomass levels, are widespread. To account for phytoplankton biomass fluctuation, a set of eutrophic lakes (167 observations from 31 systems) were included in this study (maximum chla 68 mg m^-3). The BNNs were evaluated through five assessments including single day and time series match-ups with OLCI and MSI. OLCI BNN accuracy gains of >25% and MSI BNN accuracy gains of >15% were achieved in the assessments when compared to chla reference algorithms for oligotrophic waters (chla ≤ 8 mg m^-3). In comparison to the reference algorithms, the accuracy gains of the BNNs decreased as chla and trophic levels increased. To measure the quality of the provided BNN uncertainty estimate, we calculated the prediction interval coverage probability (PICP), Sharpness and mean absolute calibration difference (MACD) metrics. The associated BNN chla uncertainty estimate included the reference in situ chla values for most observations (PICP ≥ 75%) across the different performance assessments. Further analysis showed that the BNN chla uncertainty estimate was not constantly well-calibrated across different evaluation strategies (Sharpness 1.7-6, MACD 0.04-0.25). BNN uncertainties were used to test two chla improvement strategies: 1) identifying and filtering uncertain chla estimates using scene-specific thresholds, and 2) selecting the most accurate prior atmospheric correction algorithm per individual satellite observation to retain chla with the lowest BNN uncertainty. Both strategies increased the quality of the chla result and demonstrated the significance of uncertainty estimation. This study serves as research on Bayesian machine learning for the estimation and visualisation of chla and associated retrieval uncertainty to develop harmonised products across OLCI and MSI for small and large oligo- and mesotrophic lakes.