Trophic status estimation of case-2 water bodies of the Godavari River basin using satellite imagery and artificial neural network (ANN)

Trophic status classification of inland waterbodies, using ecological indicators such as chlorophyll, turbidity, coloured dissolved organic matter (CDOM), total phosphorus, total nitrogen, and Secchi disk depth, allows ecologists to assess the overall health of an ecosystem (Matthews et al. 2012; Dodds & Smith 2016; Gholizadeh et al. 2016; Lucas et al. 2022; Muhoyi et al. 2022; Sherjah et al. 2022, 2023). The optically sensitive chlorophyll concentration is a proxy indicator of estimating the trophic status of water bodies as it is directly proportional to algae growth (Kudari & Kanamadi 2008; Srichandan et al. 2019). According to previous research outputs, untreated wastewater from industries, agricultural runoff, and municipal waste discharged into surface water bodies increases nutrients in the water, such as nitrates and phosphates, increasing algae development, known as an algal bloom (Smith et al. 1999; Tiwari et al. 2017). Subsequently, it causes a shift in the trophic status of the water bodies. The influence of anthropogenic activities on the eutrophic frequency and eutrophication risk in the tributaries of major river basins can accelerate the evolution of trophic status (Nagdali & Gupta 2002; Alongi et al. 2003). The knowledge of spatial and temporal dynamics of these trophic impacts facilitates identifying the point and non-point pollution sources. The advances in satellite remote sensing pave the way towards achieving these tasks (Kutser 2009).

Satellite remote sensing water quality methods overcome the critical barriers to traditional surface water quality, such as cost, inaccessibility, time, and effort in water quality monitoring, especially for inland water bodies which are optically more complex (Gholizadeh et al. 2016; Vakili & Amanollahi 2020). Remote sensing of water quality of inland water bodies, designated in case-2 type, becomes more complex using coarser spatial and spectral resolution sensors. Though many researchers studied chlorophyll concentration in case-1 and case-2 water using remote sensing techniques, the precise prediction remains challenging (Ruddick et al. 2001; Wang & Yang 2019). Because of the influence of adjacent optically sensitive parameters such as suspended, dissolved solids, and CDOM, connection with specific reflectance becomes minimal in the extraction of chlorophyll. The advances in satellite remote sensing technology in terms of its sensor resolution become prospective in monitoring of case-2 water bodies.

The European Space Agency's (ESA) Copernicus program, launched in 2015, allows the use of Sentinel satellite data types, publicly available optical remote sensing data for the research community (Drusch et al. 2012). Sentinel-2 acquires images with higher spatial resolutions, more spectral bands, less revisit time, and a broader swath than the Landsat satellite series. It has potential in various fields, such as land cover and land use, aquatic science, agricultural applications, and soil research (van der Meer et al. 2014; He & Mostovoy 2019). The potential of Sentinel-2 datasets for assessing the optically sensitive water quality parameters was found to be effective through the spectral band values and indices (Sent et al. 2021). The red edge bands of Sentinel-2 made it possible to calculate the indices related to the algal bloom occurrence compared to Landsat-8 (30 m) for continuous monitoring. Also, Sentinel-3 datasets with Ocean and Land Colour Instrument (OLCI) of 300 m spatial resolution and frequent revisit time (1–2 days) were prospective to generate chlorophyll-a products for continuous global monitoring (Wang et al. 2020; Bramich et al. 2021).

The computational techniques applied for connecting the satellite-derived parameters with trophic status through chlorophyll-a estimation were mainly restricted to band ratios and indices-based approaches (Du et al. 2016; Buma & Lee 2020; Chu et al. 2021). These methods were constrained to identify the high biomass waters with intense eutrophication levels. Case-2 water quality monitoring using remote sensing is complex due to its diverse and dynamic ecosystems. The research in Indian River basins is often restricted to coarser resolution satellites applied with band ratios or indices approaches (Prasad et al. 2020). The fluorescence line height (FLH) model was found to be more effective than band ratios and indices for low levels of biomass waters (Buma & Lee 2020). The applications of advanced computational techniques like ANN using satellite data synergy are in their infancy, especially in Indian river basins. The availability of multi-temporal datasets from various sensor platforms and their potential to extract trophic levels of case-2 water bodies were the main impetus of the present study. Based on these aspects, the present study aims to develop approaches for estimating the trophic status of the Godavari River basin of India using the combinations of Sentinel-2 and Sentinel-3 datasets. The study's specific objectives include:

• 1.

  Comparing Landsat-8 (OLI) and Sentinel-2 (MSI) sensors for chlorophyll-a estimations using the empirical approaches of bands and indices.

• 2.

  Identifying the feasibility of the FLH method for chlorophyll-a estimation using Sentinel-2 datasets.

• 3.

  The primary focus of the study was to develop the ANN architecture for chlorophyll-a estimation using Sentinel-2 and Sentinel-3 datasets.

Water samples from the Godavari River and its tributaries were taken from 3 March to 30 March 2022. Multiple locations (shown in Figure 1) throughout the basin, including flowing rivers from the west to south, and a partially northeast portion of the basin were also visited. The maximum number of stations were in the Telangana region, while only a few were in the boundary region between Telangana and Andhra Pradesh. The stations were selected based on the stream network, proximity, and pollution load data from Central Pollution Control Board (CPCB) and State Pollution Control Board (SPCB). Depending on the basin's geological conditions and land use, land cover stations were selected and sampled in almost 12 out of the 15 visited locations. At each sampling location, surface water was tapped by immersion of a 1 litre sample bottle from the bridge using a bailer or simply by hand on the river bank. In addition, water was collected at a depth of 1 m from the top surface for chlorophyll. The instrument immediately logged the depth reading concerning chlorophyll concentration, and a depth level gauge was used to double-check the depth. The field measurements were collected with the help of the in situ water quality testing instrument named Aqua TROLL 500 Multiparameter Sonde, which was used along with the chlorophyll-a sensor to test water quality at the sampling sites/locations. These observations were synchronized with satellite data acquisition dates to achieve good results for optical parameters. A handheld GPS was used to record the latitude and longitude of the particular station for mapping the sampled stations.

For each of the 12 sample stations, all Landsat-8 and Sentinel-2 satellite bands were extracted in GEE. Sentinel-3 images were extracted from the European organization for the exploitation of meteorological satellites. An intergovernmental European operational satellite agency (EUMETSAT) keeps tracking weather, climate, and the environment from space. The data were retrieved based on the field collection date, and considering the cloud cover percentage of less than 10%. The data were downloaded in the form of the Sentinel 3A Level-1 product Earth Full Resolution (EFR), which represents Top of Atmosphere (TOA) and is available through either the European Space Agency (ESA) SciHUB or the website: https://coda.eumetsat.int/.

The flowchart shows the overall methodology adopted in the present study (Figure 2).

The band-specific details of Landsat-8 (OLI) and Sentinel-2 (MSI) are shown in Table 1. The surface reflectance values of various bands concurrent to the sample locations were extracted in GEE and compared with in situ chlorophyll-a values and Sentinel-3 estimated chlorophyll-a values. Sentinel-3 chlorophyll data were extracted for the Godavari River basin to compare with on-site observations. To achieve this task, the data were processed in the Sentinel Application Platform (SNAP) software which is exclusively developed for Sentinel data processing. The step-by-step procedure to retrieve the chlorophyll data from Sentinel-3 (OLCI) for case-2 water involves two main steps: (1) applying visible atmospheric channel correction to obtain water-leaving radiance and (2) applying the bio-optical algorithm to retrieve the water quality.

Atmospheric correction is the process of retrieving water reflectance from total reflectance at the sensor. Then, the retrieved spectral radiance is used to determine the chlorophyll content using any of the appropriate methods. In this study, the Case 2 Regional Coast Colour Processor (C2RCC Processor) algorithm was used for Sentinel-3 (OLCI) to retrieve chlorophyll concentration. The entire process of atmospheric correction was done in the SNAP software. Following image processing in the SNAP platform, a specific area of interest was selected for further study using an image subset. The C2RCC Processor was applied to portray atmospheric conditions accurately. The data were represented using a pseudo-colour range, allowing for evaluating different chlorophyll data ranges. The image was then projected onto the proper zone, which was World Geodetic System (WGS) 84/Universal Transverse Mercator (UTM) zone 44N for this investigation. The files were then converted to raster format and used in QGIS to classify the trophic level of the basin.

The kNDVI was applied in the present study to overcome the saturation of NDVI values at higher biomass levels of water bodies following the concept applied in the case of terrestrial vegetation (Camps-Valls et al. 2021).

An artificial neural network (ANN) model was used in this study to predict the chlorophyll concentration. A three-layered ANN model consists of an input layer, a hidden layer, and an output layer. The input and output layers were connected with the neurons, and each connection had a specific weight, which was determined during the training process of the neural network. The hidden and output layers also included a bias node, which enabled the generalization of the network's classification beyond the origin. The weights were initialized randomly and adjusted iteratively to achieve optimal weights after the training. The neural network's training was done using the robust backpropagation learning algorithm, which updated the weights based on the error between the predicted and actual output. The data used for training and testing the ANN model were pre-processed (normalized from −1 to 1) to ensure that it met the conditions for successful ANN training. In this study, a single-layer neural network was employed to forecast the chlorophyll concentration at a specific sampling station by utilizing the Sentinel-2 band reflectance values of the corresponding chlorophyll sample station. For developing the architecture for large volumes of data, Sentinel-3-derived chlorophyll-a values were considered as outputs. The number of hidden nodes in the network was adjusted based on the performance of the training algorithm, as detailed in the results section. The data were split into a training set and a test set, with the training set being used to train the ANN model and the test set being used to evaluate the performance of the trained ANN model using the regression coefficients.

The section briefly discusses the results obtained by correlating in situ chlorophyll-a measurements with satellite-derived observations of Landsat-8 and Sentinel-2 datasets. The main focus of the section is to identify the appropriate satellite sensor platform out of Landsat-8 (OLI) and Sentinel-2 (MSI) datasets for correlating with in situ chlorophyll-a observations. Furthermore, suitable bands sensitive to chlorophyll-a variations were identified from the selected satellite datasets.

We also experimented with various band combinations to identify the existence of a better relationship with chlorophyll-a concentration. In this regard, the two-band difference approach between SWIR-1 and SWIR-2 disclosed better relation than all other possible combinations (Figure 3). The physical aspect of this relationship can be attributed to the characteristics of these channels to differentiate between clear waterbodies and turbid regions. The difference between the SWIR-1 channel and the SWIR-2 channel is directly proportionate to the volume of phytoplankton in the water bodies, which is evident in our results. Apart from different possible combinations of bands, we also tested the scope of various greenness and water content-related indices to relate with chlorophyll-a variations. In this context, we checked the suitability of NDWI, MNDWI, NDVI, and kNDVI for connecting with field-measured chlorophyll-a values.

Furthermore, to evaluate the efficacy of red edge spectral reflectance in double band models of the Sentinel-2 satellite imagery, the performance of adjacent band difference relationships with chlorophyll-a was also checked. Similar to the results obtained in the case of Landsat-8, the SWIR adjacent channel's difference (B11–B12) also showed better performance but was not superior to the red edge and NIR bands. The results showed a significant correlation (p < 0.01) by adjacent red edge band differences ((B5–B6) and (B6–B7)) as well as the NIR region bands (B8–B8A). These outcomes correlate with the fundamentals of red edge area changes proportionate to the variations of phytoplankton volume. These facts may facilitate the development of a unique water quality index sensitive to trophic status dynamics of case-2 waterbodies.

The superior performance of the green, red, and NIR regions (R² = 0.91; RMSE = 1.6 μg/l) correlates with their sensitivity concerning the fluorescence variations within the visible NIR spectrum. These findings also connect the output obtained for the empirical approach of kNDVI, which includes the bands of red (B4) and NIR (B8) as inputs. The spatial resolution of 10 m of these band combinations also benefits fine resolution mapping compared to other band combinations.

The outcomes obtained until the sections were constrained to the measured in situ field-measured chlorophyll-a values. The scope of wide ranges of chlorophyll-a values from field samples is limited, considering the requirement of concurrent satellite passes and cloud-free data availability. Hence, it is essential to check the performance of these methods for a wide range of values of trophic status from alternate sources of chlorophyll-a values. The further section verifies the possibility of applying the developed FLH for broader ranges extracted from Sentinel-3 datasets.

The plot shows a strong correlation between the in situ chlorophyll-a values and Sentinel-3 (OLCI) derived values with a slight overestimation of values by Sentinel-3 derived outputs. The coarser spatial resolution of Sentinel-3 data of 300 m and the optical complexity of case-2 water bodies induce variations from actual values of chlorophyll-a concentration. The influences of these factors are minimal while focusing on the classification of trophic status as it spans over ranges than the absolute values. Based on these assumptions, we have utilized the Sentinel-3 data-derived values for trophic status estimations.

The scheme adopted for trophic state classification in the present study is presented in Table 2 (Carlson 1977). The chlorophyll-a estimation is considered a proxy parameter for trophic status estimation due to its strong association with water quality and also considering its optical characteristics in the remote sensing perspective.

The map shows varying levels of chlorophyll-a concentration spanning from 0.06 to 98 μg/l, with most water bodies falling under the range of 2.6 to 55 μg/l. The trophic level classification of the generated map was carried out as per the selected scheme, and the four trophic level-based maps are presented in Figure 11.

The distributions of trophic status showed mesotrophic status for 40% of the basin. The eutrophic status also contributed to 28% of the case-2 water bodies of the region, which needs further attention in these regions. The lower part of the basin was more eutrophic and most of the water bodies in these regions were covered with agricultural fields, which was noticed during the field visit. The agricultural practices followed may be the reason for the dominant eutrophic status in these regions. These results indicate the need to correlate the trophic status with the land use/land cover map of the region to reveal the specific reasons for eutrophic conditions.

A feedforward neural network with one input layer, one hidden layer, and one output layer was simulated using MATLAB for chlorophyll-a estimation. The Sentinel-2 band values were considered as inputs for the study and appropriate bands were selected based on the correlation between the values. The redundant bands were eliminated and only the unique bands were considered (from Figure 13). The number of hidden nodes was chosen through trial and error, and the maximum and minimum hidden nodes were tried from 0 to 15 using the 2n + 1 thumb rule given by the Kolmogorov mapping theorem (Hecht-Nielsen 1988) where n is the number of inputs.

The outcomes reveal the potential of ANN developed using the Levenberg–Marquardt algorithm with Sentinel-2 band values as inputs for estimating chlorophyll-a for higher ranges and the feasibility to estimate the eutrophic and hyper-eutrophic status conditions. Further investigations on implementing the developed architecture in the GEE platform may enable an understanding of time-series analysis of trophic status using Sentinel-2 datasets at a finer resolution of up to 10 m. Hence, the knowledge of the near real-time trophic status of case-2 water bodies can be generated for different climatic conditions and show the scope of generating operational products for trophic status estimations.

The preliminary investigations related to the choice between Sentinel-2 and Landsat-8 for trophic status estimation studies of case-2 water bodies showed the superior performance of Sentinel-2 (MSI). The role of the NIR spectral channel as a single-band index for chlorophyll-a estimation was evident for both cases. The critical result obtained in the present study is the capability showcased by kNDVI for chlorophyll-a estimations compared to single bands and other selected indices. The green, red, and NIR regions were found to be appropriate channels for FLH estimation and could accurately estimate chlorophyll-a up to 20 g/l. The trophic status of the Godavari basin estimated using Sentinel-3 data-derived chlorophyll-a values revealed the existence of 28% eutrophic status waterbodies within the basin. These results prompted the development of ANN architecture using the Levenberg–Marquardt algorithm. The performance of the ANN approach showcased superior performance for mapping all levels of the trophic status of case-2 water bodies. The Sentinel-2 data-driven architecture, supported by Sentinel-3 chlorophyll-a as outputs, showed a remarkable potential for estimating the near real-time trophic status for all levels of trophic status with significant accuracy.

The investigations using the large volume of in situ observations may disclose its further potential for higher biomass mapping, which needed to be clarified due to the lack of enough data in this study. Also, there is a scope for utilizing new datasets like Landsat-9, which has 14 bits radiometric resolution and provides better sensitivity to water-leaving radiance values. Due to these enhanced resolutions, it may have more sensitivity to chlorophyll-a concentration compared to Landsat-8 and Sentinel-2, which we applied in this study.

The authors would like to thank the Council of Scientific and Industrial Research (CSIR), India grant (No. 24(0356)/19/EMR-II) of the project titled ‘Experimental and Computational Studies of Surface Water Quality Parameters from Morphometry and Spectral Characteristics’.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.