There are ~100 million lakes in the world at a wide range of spatial scales. For large lakes we can derive water quality parameters using ocean colour satellite sensors, which benefit from a long history of algorithm development and validation, and dedicated bio-optical data collection. For the optically complex waters found in lakes a rich body of work has been dedicated to developing algorithms for the MERIS sensor, and the follow-on mission OLCI benefits from this legacy. Due to the optical diversity of inland waters, algorithms intended for global use require in situ validation data from a wide range of optical water types. For the MERIS and OLCI sensors their long time series means a diverse selection of in situ data are more readily available.
The 300m resolution of MERIS/OLCI imposes a lower size limit on the lakes that we can monitor with these sensors, which has led researchers to look to the MSI sensor on Sentinel 2 as a higher resolution alternative. MSI was primarily designed for land remote sensing so lacks the radiometric resolution and sensitivity of the dedicated ocean colour sensors. Furthermore, the relatively short time series limits the availability of in situ data for validation. Currently the available in situ data for MSI does not adequately represent the diversity and seasonality of optical properties of lakes to calibrate and validate algorithms to be used on the global scale.
While our ultimate goal is to improve algorithm performance for MSI using in situ data, we can make improvements in the short term by comparing with established satellite sensors and algorithms. Here, we present work to align the results of turbidity and chlorophyll-a algorithms for MSI against concurrent OLCI data, resulting in significant performance improvements compared with the original algorithm coefficients.
In this presentation we will outline our approach and show results from the tuned algorithms. The remaining challenge is to explore how well these aligned algorithms perform when applied to smaller lakes where issues such as land adjacency affect a large proportion of the water body. We will present examples where the aligned algorithms perform well and where they fail. As the output of these algorithms are available at 100-m resolution in the Copernicus Global Land Service Lake Water Quality product, we will make recommendations on the type of applications and water bodies for which the current procedures are adequate, and on which priorities we see for the further development of high-resolution inland water quality products.
Remote sensing reflectance (Rrs) is the conventional measurement used in aquatic remote sensing to spectrally characterize optically active constituents (OACs) just below the water surface and is often the desired parameter to be used in empirical and analytical bio-optical models for water quality applications. Deriving Rrs from Earth observing multi-spectral image pixels is a major challenge as the water-leaving reflectance is only a fraction of the total signal received by a space-borne sensor due to the interfering scattering and absorbing properties in the atmosphere. Though several algorithms and software have been developed to alleviate this process through atmospheric correction for global ocean color sensors, it is still a very active area of research for medium resolution sensors for inland water remote sensing. To address this concern, the U.S. Geological Survey (USGS) released the Landsat-8 Collection 1 Level-2 Provisional Aquatic Reflectance (AR) Science Product in April 2020 after the Operational Land Imager (OLI) sensor showed a high degree of fidelity to derive reliable Rrs measurements at 30 meters resolution over coastal and inland validation platforms and coincident observations with other ocean color sensors.
The AR science product is based on the Sea-viewing Wide Field-of-View Sensor (SeaWiFS) Data Analysis System (SeaDAS) originally developed by the National Aeronautics and Space Administration Ocean Biology Processing Group (NASA/OBPG) modified for the Landsat-8 OLI sensor. The SeaDAS algorithm processes Landsat-8 Level-1 TOA reflectance bands to estimate Rrs through precomputed radiative transfer simulations that depend only on the sensor spectral response, solar and sensor viewing geometry, and ancillary information such as atmospheric gas concentrations, surface windspeeds, and surface pressure. Spectral Rrs are then normalized by the Bidirectional Reflectance Distribution Function (BRDF) of a perfectly reflecting Lambertian surface (multiplied by π) to produce the dimensionless Aquatic Reflectance. Global AR products for Landsat-8 Collection 1 and Collection 2 data are made available for download through the Earth Resources Observation and Science (EROS) Center Science Processing Architecture (ESPA) on-demand interface to help support ongoing contributions to aquatic remote sensing and environmental monitoring capabilities.
With scheduled release of Landsat Collection 2 AR provisional products in early 2022, improvements were made to the water pixel classification for non-water masking that will enhance the capability to evaluate spectra from AR products across a range of optically diverse targets. For qualitative uncertainty in aquatic reflectance spectra, SeaDAS generates a Level-2 processing flags (l2_flags) band that provides additional per-pixel information about the Landsat-derived spectra. Additionally, intermediate Rayleigh-corrected reflectance (ρrc) products for the visible to shortwave-infrared (VSWIR) bands will be included in the Collection 2 AR product package as well as the original atmospheric auxiliary input raster bands used in the SeaDAS processing for advanced pixel analysis.
As new Earth observing satellite sensors become operational as part of long-term continuity missions designed and launched by NASA (National Aeronautics and Space Administration), there will be a heightened need for data synergy and cross-sensor harmonization to provide increased observational frequency that yield comparable measurements. With the recent launch of Landsat-9 OLI-2 as well as proposed plans for future Landsat missions, the AR science product is a first-step toward structuring a standardized unit of measurement for the health and changing dynamics of aquatic ecosystems.
This presentation will provide the aquatic remote sensing community with a description and general use of the Landsat Provisional Aquatic Reflectance Science Product, its characteristics, product packaging, download accessibility, and future implementation to facilitate its continued use in water management practices and global water surveying.
Soda lakes in the East African Rift Valley are some of the most productive aquatic ecosystems on earth. These lakes sustain more than half of the global population of Lesser Flamingos, which feed on dense blooms of the cyanobacteria Arthrospira fusiformis and microphytobenthos. Given their dependence on a small network of Rift Valley lakes, Lesser Flamingos are highly vulnerable to both climate change and catchment degradation. This study aimed to investigate: i) the extent to which the water quality of soda lakes is changing; ii) what is driving these changes - e.g., land-use or climate change; and iii) how these water quality changes are affecting the numbers and distribution of Lesser Flamingos.
We used Landsat 7 ETM+ to examine the water quality trends of 14 Rift Valley lakes over a 21-year time period. Low cloud images were selected, and monthly composites were generated using the median reflectance for each pixel. A decision tree-based classification algorithm, originally developed by Tebbs et al. (2015) for Rift Valley lakes, was adapted to include salt crust as an additional class. The classification scheme utilises the distinct spectral signatures of different ecologically relevant optical water types to group pixels into pre-defined classes. The final classification included two classes that represent valuable food sources for flamingos (high biomass waters and microphytobenthos) and five further classes (low biomass, sediment dominated, surface scum, bleached scum and salt crust). The method was then applied to the monthly Landsat ETM+ composites to estimate lake ecological status and flamingo food availability.
Environmental parameters were also estimated from satellite imagery to assess the impacts of land-use and climate change on lake ecological states. Using Landsat 7 ETM+ imagery, the Modified Normalised Difference Water Index (MNDWI) was applied to estimate lake water levels and water surface temperature was estimated using the statistical mono-window (SMW) algorithm applied to Top of Atmosphere (TOA) imagery. The Normalised Difference Vegetation Index (NDVI) was applied to MODIS imagery to assess land degradation within catchments and precipitation was estimated using the Climate Hazards Group Infrared Precipitation with Station (CHIRPS) data. All remote sensing analyses were performed in Google Earth Engine. Ground-counts for Lesser Flamingo numbers were obtained from the International Waterbird Census. Water quality (as an indicator of food availability) and flamingo distributions were then modelled using generalised linear mixed models.
Results confirmed that Lake Bogoria and Lake Nakuru in Kenya are important feeding lakes for Lesser Flamingos. Food resources in these lakes have been declining in recent years, without increasing sufficiently at other lakes. Declines in food availability were linked to rising water levels, which are occurring due to a combination of increased rainfall and land degradation within the lake catchments. Food availability was a key driver of flamingo distributions, with water levels also having an influence. These findings have important implications for conservation and land and water resource management within the East African Rift Valley. This study also demonstrates the potential of optical classification of inland waters using Landsat imagery for the long-term monitoring of lake water quality and ecological status.
Water quality is one of the major issues that societies are already facing and there is a strong need of a global monitoring on water quality in inland waters. Satellite data are increasingly viewed as a trustworthy solution to complete the ground observations that are severely lacking in most parts of the world.
Research teams from the IRD (GET Laboratory), OFB and INRAE (ECLA research group) research institutions, supported by the French Space Agency (CNES) have developed the OBS2CO processing chain for inland water color analysis and water quality mapping based on Sentinel-2 images which resolution makes possible to assess very small objects, of a few tenths of meters, such as rivers and minor lakes. The processing chain handles Sentinel-2 Level-1C products and takes advantage of two modules specifically designed for complex optical water and small objects : 1) Glint Removal for Sentinel-2 images that operates both atmospheric correction and sunglint correction (Harmel et al. 2018); 2) Water Detect module that retrieve water pixels using unsupervised images and proved to be very efficient for small water objects, down to 1 ha (Cordeiro et al. 2021). Both codes were tested during large intercomparisons tests (ACIX-AQUA for atmospheric correction over optically complex waters and WorldWater Round Robin for water surface detection) and proved to provide robust results. The OBS2CO chain delivers water occurrence maps as well as water quality maps for different parameters : suspended particulate matter (SPM), turbidity, chlorophyll-a concentration, Harmfull Algal Blooms index, Coloured Dissolved Organic Matter (CDOM) absorption coefficient at 440 nm. A large database of water quality and field radiometric measurements (i.e. > 1000 points) collected in river and lakes across the world is used to evaluate the retrieval algorithms.
The first water color products from Sentinel-2 images are distributed through UNESCO World Water Quality Portal over Lake Chad Region in Africa (https://lakechad.waterqualitymonitor.unesco.org/portal/) and will be shortly available at the French land datacenter. We will explain the method, calibration of the model and its validation over various sites worldwide and present the freely distributed water color products.
The Sentinel-2 high resolution makes possible to retrieve very fine details related to hydrological and biophysical processes in rivers and lakes. Different applications of the OBS2CO processing chain will be presented over Europe, South America, Africa and Asia. These applications consider suspended sediment flux quantification over rivers and reservoirs, eutrophication monitoring in semi arid areas and for EU water framework directive assessment, dissolved organic matter mapping in tropical areas and disaster assessment.
Invasive floating aquatic plants, such as Water Hyacinth (Pontederia crassipes) and Water Lettuce (Pistia stratiotes), have substantial negative ecological and socio-economic impacts, affecting lake water quality and impeding fishing and navigation. Due to their free-floating nature and rapid proliferation, they pose a significant logistical issue in their removal and require continuous ground-based monitoring, which is both expensive and time consuming. Existing methods for remote sensing of floating plants have limited geographic or temporal coverage and are not easily transferable to other open water systems. The main limitations were the spatial, temporal or spectral characteristics of past sensors, now mostly resolved by the recent Copernicus Sentinel-1 and Sentinel-2 missions.
In this presentation, we propose generalised and well-validated methods for satellite remote sensing of floating aquatic plants using radar (Sentinel-1 SAR) or optical (Sentinel-2 MSI) imagery and evaluate the strengths, limitations and complementarity of radar and optical imagery. The spectral properties of floating macrophytes are investigated, and we evaluate spectral indices suited to shoreline delineation and distinguishing floating plants from surface algae and open water. The methods developed were tested on eight diverse waterbodies distributed globally across three continents. We discuss the challenges of ground-truthing floating macrophyte maps due to the dynamic nature of free-floating plants and present a standard protocol for generating validation data from available satellite imagery.
Owing to the similar spectral and backscatter properties of land and floating plants, multi-temporal imagery was used for shoreline delineation. The Automated Water Extraction Index (AWEI) and Normalised Difference Moisture Index (NDMI) applied to Sentinel-2 optical imagery were used to distinguish floating plants, surface algal blooms and open water, and radar imagery from Sentinel-1 VH backscatter was used to distinguish floating plants from open water. Methods were able to map floating plants with user’s accuracies of 86.4% (Sentinel-1) and 79.8 % (Sentinel-2) and producer’s accuracies of 84.3% (Sentinel-1) and 80.4% (Sentinel-2). Timeseries for the eight waterbodies were produced showing the change in floating plant coverage from December 2018 – June 2021.
The strengths of Sentinel-1 radar included being unimpeded by cloud cover and floating surface algae. It was, however, influenced by false positives due to wind and thermal noise artifacts and was less able to resolve narrow river channels. The two satellites provided complementary information on floating weed location and coverage, with significant agreement on same-day measures of floating weed area (R2 = 0.85, p < 0.05). We therefore suggest the benefits of using both sensors together for detection and monitoring of floating macrophytes. Our novel methodological framework provides a valuable tool for researchers and water managers for monitoring the spread of water hyacinth and other floating weeds to target control measures more cost-effectively.
Cyanobacteria occurrences are more and more frequent in many water bodies around the world. Due to their potential harmfulness, the detection and monitoring of cyanobacteria blooms are especially relevant in freshwater bodies for communities with implanted recreational activities as well as for government instances assessing and monitoring water quality for reporting purposes.
Various studies have shown that satellite data can be used to quantify cyanobacteria blooms in lakes and coastal waters. These approaches include the assessment of floating cyanobacteria or the detection of phycocyanin in cyanobacteria blooms. Using remote sensing to detect cyanobacteria in the water column is based on the spectral behavior of phycocyanin, which, in contrast to chlorophyll-a, has a clearer absorption maximum at 620nm. At the same time there is an absorption minimum at 650 nm and at 700 nm. To be able to detect these absorption features with an optical sensor and to distinguish them from chlorophyll-a, the corresponding sensor must have narrow recording channels at the necessary wavelength ranges. The required phycocyanin absorption band at 620nm is only available for a few sensors, such as MERIS (on board ENVISAT 2002-2012) and OLCI (Sentinel-3, since 2016). The spatial resolution of both sensors is 300 m and is therefore only suitable for larger lakes. In the case of sensors with a higher spatial resolution, which are often optimized for land applications, there are usually wider recording channels and the area around 620 nm is not explicitly covered with its own band at 620 nm. Developing a method to detect cyanobacteria with high spatial resolution sensors is therefore not straightforward but essential to not only take smaller lakes into account during monitoring but also to investigate the spatial extent and behavior of cyanobacterial blooms in water bodies. Sentinel-2 MSI has proven to be a valuable sensor for monitoring smaller (inland) waters bodies but is missing the 620nm band. Nonetheless, analyzing these inland water bodies with Sentinel-2 MSI showed that high cyanobacteria abundances present distinctive features in the spectra and the water colour which can be used to derive an indicator for potential risk of cyanobacteria occurrences, by e.g. using a B5/B4 band ratio.
We present in this approach a random forest (RF) model as a regressor to identify the risk of cyanobacteria blooms with Sentinel-2. RF is a regressor that randomly produces multiple decision trees, with each tree representing a class prediction. The class with the most votes becomes the model’s prediction. The RF method is used for regression or classification approaches, and it belongs to the supervised machine learning algorithms. The RF method uses the wisdom-of-crowds concept and is becoming popular in remote sensing applications thanks to easy and fast model training. The importance of the used variables has been extensively tested for different scenarios. A manually selected training dataset with known cyanobacteria blooms, high chlorophyll biomass blooms and clear water cases has been collected covering various inland waters from Germany and the USA. Since atmospheric corrections may fail in extreme blooming events or generate higher uncertainties, we decided to use bottom of Rayleigh reflectances (BRR) spectra instead. This has also the advantage to be independent on a certain atmospheric correction algorithm. Four spectral indices were determined to represent the occurrence of cyanobacteria in inland waters. During the validation phase challenges for the model were identified in brown waters with cyanobacteria occurrences. These challenges were resolved by introducing a second model trained with a subset of the entire training’s dataset. The identification of cyanobacteria is finally based on the combination of these two models. The output of the combined model is a risk status of potential cyanobacteria occurrences. The risk status is subdivided into low, medium, and high risk.
The validation of the developed approach was performed on German lakes, based on an extensive in-situ dataset from the German state offices. For the validation dataset, Cyanobacteria occurrences from in-situ data were classified if the biovolume of cyanobacteria were higher than half of the total biovolume. Additionally, a minimum chlorophyll-a concentration of 10µg/L needed to be recorded to account for a cyanobacteria bloom. Furthermore, a selection of lakes in the US based on a database of cyanobacteria identification compiled by the NRDC (Natural Resource Defence Council, 2019) were analysed. The database covers events including any type of response that may be related to a harmful algal bloom (HAB) including a cyanotoxin detection or reported illness. With an overall accuracy of 0.88 using German lakes, the validation shows promising results to identify cyanobacteria blooms in inland waters. Further analyses are ongoing to test the algorithm related to high biomass blooms falsely detected as cyanobacteria risks and the cyanobacteria identification within rivers.
The strength of the presented approach is in the implementation of a comprehensive trained random forest model, independent from atmospheric correction uncertainties, covering various cyanobacteria conditions in inland waters. However, some challenges are still present in the model approach, such as uncertainties for high biomass bloom without cyanobacteria or the missing information about the cyanobacteria concentration. While the approach is under development, if proven robust, the setup would allow for an expansion towards a risk assessment of cyanobacteria in inland waters for the high-resolution sensor Sentinel 2 MSI.