Retrieval and mapping of chlorophyll-a concentration from Sentinel-2 images in an urban river in the semiarid region of Brazil

Recently, the Poti river mouth region has experienced environmental impacts that resulted in a change of landscape in its dry season, highlighting the eutrophication and proliferation of phytoplankton, algae, cyanobacteria and aquatic plants. Considering the aspects related to water-quality monitoring in the semiarid region of Brazil from remote sensing, this study aimed to evaluate the performance of Sentinel-2A satellite data in the retrieval of chlorophyll-a concentration in Poti River in Teresina, Piaui, Brazil. The chlorophyll-a concentration retrieval and mapping methodology involved the study of the water surface reflectance in Sentinel-2A images and their correlation with the chlorophyll-a data collected in situ during the years 2016 and 2017. The results generated by the Chl-1, Ha et al. (2017), Chl-2, Page et al. (2018), and Chl-3, Kuhn et al. (2019) equations show the need for calibrating the algorithms used for the Poti River water components. However, the empirical algorithm Chl-2 shows a correlation has been established to identify the spatiotemporal variation of chlorophyll-a concentration along the Poti River broadly and not punctually. The spatial distribution of this pigment in maps derived from Sentinel-2A is consistent with the pattern of occurrence determined by the in situ data. Therefore, the MSI sensor proved to be a tool suitable for the retrieval and monitoring of chlorophyll-a concentration along the Poti River.


INTRODUCTION
Disorderly development in large urban centres, devoid of any planning and with increasing levels of environmental degradation, drastically affects water availability and especially its quality (Vargas et al., 2018). The quality of the river water is affected by factors such as the increase of domestic and industrial pollutants that favour the eutrophication of the river water. The eutrophication that can lead to the death of aquatic fauna and flora, turning the water unfit for consumption and other uses (Esteves, 2011;Jorge and Lobo, 2019).
Water-quality monitoring is a challenging process as data collection is insufficient or nonexistent for most bodies of water. Punctual samples do not always portray the dynamics of water constituents, because the river is a highly dynamic lotic ecosystem that requires the collection of a large number of quality parameters to understand spatiotemporal variations and monitor changes (Prasad et al., 2018;Kuhn et al., 2019;Martins, 2019). In contrast, some authors have adopted an efficient approach that integrates diverse in situ data with remote sensing images to monitor the factors that affect water quality and understand the limnological processes, because satellite imagery provide the synoptic, continuous and long-term observation (Ha et al., 2017;Pinardi et al., 2018;Martins, 2019).
The European Space Agency (ESA) has launched the mission Sentinel-2 with two identical satellites. The MultiSpectral Instrument (MSI) onboard the satellites is a 13-band passive optical sensor in a 20.6° orbital field of instrumental view. These features, in addition to a 10 m spatial resolution and five-day high temporal resolution, make the MSI sensor a suitable instrument for the retrieval of biophysical parameters in inland waters compared to other satellite missions designed for such applications as Landsat-8 / OLI, Sentinel-3 / OLCI, Aqua / MODIS (ESA, 2019;Pinardi et al., 2018;Pereira-Sandoval et al., 2019).
Studies on the use of Sentinel-2A images in the retrieval and mapping of chlorophyll-a focus on a limited number of band ratio empirical algorithms. These models are based on statistical relationships between in situ measurements of water constituents and radiometric data from the satellite sensor, which provide the best correlation between reflectance data and the concentration of optically active water constituents at different wavelengths (Ogashawara et al., 2017). The commonly used two band-ratio algorithms are based on the ratios of the blue (440 and 510 nm) and green regions (550 and 555 nm), of the red (670 and 675 nm) and near Rev. Ambient. Água vol. 15 n. 2, e2488 -Taubaté 2020 infrared (NIR) regions (685 and 710 nm), and of the green (550 and 555 nm) and red regions (670 and 675 nm) (Mouw et al., 2015;Toming et al., 2016;Ha et al., 2017).
In the semiarid region of Brazil, in the city of Teresina, in the state of Piaui, the urban riverbed of the Poti River has experienced intense environmental impacts in recent years, leading to changes in its landscape in the dry season, highlighting eutrophication and the proliferation of phytoplankton, algae, cyanobacteria and mainly from aquatic plants (Costa, 2014;Santos, 2017). The Sanitation Laboratory at the Federal University of Piaui has estimated chlorophyll-a concentration in punctual samples of the Poti River from laboratory analyses in 2016 and 2017. However, this sampling is a costly and low-temporal frequency procedure. The application of remote sensing can monitor large extensions of the Poti River over a long period and with a high temporal frequency, reducing efforts and resources, because six Sentinel-2 images are acquired monthly in the study area.
In this context, this study evaluated the performance of the Sentinel-2A satellite data in the retrieval of chlorophyll-a concentrations in Poti River in Teresina, Piaui, Brazil. This is the first attempt to use remote-sensing techniques to monitor the spatial and temporal variability of water quality parameters in order to support the decision-making process, regarding preventive and corrective environmental management of the intense environmental impacts that occur in the Poti River, mainly in the dry season.

Study area
The study area, as shown in Figure 1, corresponds to the 36.8 km urban stretch along the Poti River, located in the municipality of Teresina, which is the largest city and the capital city of the state of Piaui, Brazil. This section of the river was chosen due to the location of the water quality monitoring sample points that have been carried out by the Sanitation Laboratory at the Federal University of Piaui. The Poti River originates in the state of Ceara, in the city of Algodoes, and empties into the River Parnaiba, in Teresina. River Poti's mouth flows around 1.3 m³ s-1, on average in the driest quarter. The studied area has an approximate area of 3.7 km² and an average width of 100 m in the rainy season and, during the dry season, it can be reduced to a few meters wide and a few centimetres deep in some places (Mendes-Câmara, 2011).
The region is in the transition zone between the Cerrado and Caatinga biomes, in which the predominant natural vegetation is mixed forests that are being replaced by areas of agriculture and degraded vegetation (Espindola et al., 2017). The climate of the region is characterized as tropical equatorial zone, warm, semiarid, with a dry season from July to November and a rainy season from January to April, an average annual temperature and precipitation of 27.4ºC and 1.325 mm, in that order (IBGE, 2015;INMET, 2018). In the last four decades, Teresina, with 1,391.046 km² of area, has suffered high rates of urban expansion (IBGE, 2017). The city urbanization tendencies show that urban expansion has been larger than the population growth, causing problems related to the occupation of flood regions in the confluence of the Parnaiba and Poti Rivers, the contamination of the rivers themselves, the increase of the traffic, the increase of air pollution, lack of adequate housing and general infrastructure in peripheral areas (Espindola et al., 2017). It is noteworthy that Teresina has only 61.6% of households with adequate sanitation (IBGE, 2017).

Remote sensing data acquisition and preprocessing
In this study, the remote-sensing data has been obtained by the Sentinel-2A satellite, available on the Copernicus Open Access Hub web platform (ESA, 2019). The images have the size, spatial and radiometric resolution of 100 km x 100 km, 10 m, 20 m and 60 m, and 12 bits, respectively. The Level-1C (L1C) product downloaded is pre-processed, orthorectified, georeferenced to the Universal Transverse Mercator projection using the World Geodetic System 84 datum, and radiometrically calibrated for top-of-reflectance atmosphere (ESA, 2018). The atmospheric correction was processed in the downloaded images using the Sen2Cor-2.5.5 processor, available in the Sentinel Application Platform (SNAP-6) program toolbox, resulting in Level-2A (L2A) product, radiometrically calibrated for bottom-of-atmosphere reflectance. As the algorithm used in the correction described above processes bands of different spatial resolutions, the images were resampled to 20 m by the tool (ESA, 2018). The specifications of Sentinel-2A bands used in this study are presented in Table 1 (ESA, 2018; SINERGISE, 2018).

Field Data
During 2016 and 2017, the Sanitation Laboratory sampled chlorophyll-a concentrations at seven monitoring points shown in Figure 1 monthly. The reference standards for the monitoring of bodies of water are set out in National Environmental Council Resolution No. 357 (CONAMA, 2005) and in the National Water Agency's National Sample Collection and Preservation Guide (ANA, 2011). The chlorophyll-a were obtained from water samples by the spectrophotometric analysis using a Hach DR 2800 Spectrophotometer. Samples were filtered, extracted and calculated according to colorimetric method (Marker et al., 1980). From a total of 168 samples available, 56 samples were used corresponding to the dry season. Considering that the field data sampling collecting date did not agree with the dates that Sentinel-2A overpassed the study area, we analysed all time windows available in the dry season, because it is the critical period during which eutrophication and the proliferation of phytoplankton, algae, cyanobacteria and aquatic plants occur. The number of match-ups along the period under study: 1 with one-day difference, 1 with two-days difference, 2 with threedays difference, 1 with seven-days difference, 1 with ten-days difference and 2 with twelvedays difference.

Chlorophyll-a concentration retrieval algorithms
A variety of empirical algorithms have been developed and used to retrieve chlorophyll-a concentration in inland waters, where the spectral reflectance curves of waters with different chlorophyll-a concentrations are characterized by high absorptions in the blue and red bands and high reflectance in the green and red edge bands (Matthews, 2017;Ogashawara et al., 2017). The choice of the algorithms was based on a literature review in which the band ratio empirical algorithms found in Ha et al., 2017 (Chla-1: Equation 1), Page et al., 2018 (Chla-2: Equation 2) and Kuhn et al., 2019 (Chla-3: Equation 3) were selected. These algorithms include the green to blue band ratios and those including the red band and red edge band (Ha et al., 2017;Page et al., 2018;Kuhn et al., 2019). They were routinely applied to Sentinel-2A data and obtained the best performances in retrieval chlorophyll-a in lentic and lotic inland waters from regions with similar geographical and climatic features as those found in the study area.
Discrepancies between chlorophyll-a concentrations of punctual data, in situ (Chla-is) and satellite-derived (Chl-1, Chl-2, Chl-3), were quantified using statistical metrics, often intended for evaluation of remote-sensing algorithms, which in this case include the Pearson squared correlation or coefficient of determination (R²), the Root Mean Square Error (RMSE), and the Bias (Ha et al., 2017;Page et al., 2018;Kuhn et al., 2019). All data were analyzed using R 3.6.1 and Past 3.24 statistical softwares.
The R² metric represents the linear consistency between the observations and the proportion of the variation explained by the linear regression, the RMSE measures the precision of the combinations and the absolute difference, which is sensitive to the extreme values, and the Bias determines the underestimation or overestimation of the calculated data, compared to field-measured data (Qin et al., 2017;Ansper and Alikas, 2019). A high R² value indicates a high degree of correlation between in situ observations and Sentinel-2A, while a low RMSE value indicates that Sentinel-2A observations resemble well with in situ observations, and a value close to zero for Bias suggests that there is no systematic underestimation or overestimation of Sentinel-2 and in situ data (Qin et al., 2017). Table 2 shows the descriptive statistics of in situ chlorophyll-a (Chla-is) punctual concentration levels during the sampling period. In the Poti River, spatial variability is represented by the variation in the average concentrations of Chla-is, while the high standard deviation in the concentration levels of Chla-is indicates temporal variability. Table 2. Descriptive statistics of chlorophyll-a in situ punctual concentration (mg m³-1) in 2016 and 2017. From the monthly in situ samples of chlorophyll-a, the average results obtained in 2016 were 28.67 mg m³-1, ranging from 5.46 mg m³-1, recorded at PT-00, to 76.44 mg m³-1, read at PT-02, both in August. In 2017, the average chlorophyll-a count was 10.21 mg m³-1, also varying in August at the same points, between 4.14 mg m³-1 and 33.15 mg m³-1, respectively. According to the data, annually the concentration of chlorophyll-a increases from August to December, peaking in November, and in 2017 there was a reduction of 36% in the concentration of chlorophyll-a in the study area. These results show a large spatiotemporal variation of chlorophyll-a in the dry season, depending on the location of the collection points, where the points that indicate the lowest and highest recurrent chlorophyll-a concentrations are PT-00 and PT-02.

Point / Year Mean Standard deviation Coefficient of Variation Minimum Maximum
The influence of factors such as incipient environmental management, occupation of river margins, existence of clandestine connections of raw sewage in the rain drainage and the high evaporation of water, have contributed to the alteration of the limnological characteristics of the Poti River (Oliveira and Silva, 2014;Oliveira Filho and Lima Neto, 2018). In addition, in the dry season, the reduction in flow, width and depth favour the formation of natural barriers, such as meandering curves, rocky outcrops, alluvial deposits of pebbles to sands and river islands, cause the damming of Poti River waters (Lima and Augustin, 2014). From this perspective, natural and artificial barriers, in addition to the land-use change, are the most important structural determinants for the modification of the limnological characteristics of a watershed, as they create a lentic mosaic macrosystem, very different from the original lotic condition, favourable to the processes of eutrophication (Debastiani Júnior et al., 2016). These concentrations can be explained by the location of these points, as PT-00 is located in the leastinhabited region, with preserved riparian forest and without contributions from domestic and industrial effluents, while PT-02 is located at 850 m downstream from the Alegria Sewage Treatment Plant (Mendes-Câmara, 2011).
Subsequently, the processing of the Chl-1, Chl-2 and Chl-3 equations was performed on Sentinel-2A / L2A images obtaining the values of chlorophyll-a at seven points.
Considering that these algorithms also tend to vary in performance, because depending on the optical properties of the river and chlorophyll-a, temperature, nutrients and light may introduce classification errors, the performance of the applied algorithms was analyzed punctually in Sentinel-2 images available in the dry period. In total, four values of chlorophylla were used for each sampling point. The performances of the algorithms for the Poti River are summarized in Table 3.
Considering a moderate correlation, with R² values equal or higher than 0.50, between the results of punctual quantitative agreement between in situ data and values obtained with each equation, overall, the empirical algorithm Chl-2, with seven values, obtained the best point performance compared to the algorithms Chl-1, with six values, and Chl-3, with four values. Therefore, according to this data, the relation between the bands B04 and B05 most indicated the retrieval of chlorophyll-a in the Poti River.
In this sense, considering the Equation Chl-2, by Page et al. (2018), Figure 2 shows the graphs of the dynamics of the chlorophyll-a in situ (Chla-is) and satellite-derived (Chl-2) for 2016 (samples 1-28) and 2017 (samples 29-56). In the set of 56 samples, the Chl-2 algorithm shows an overestimation of concentrations, as indicated by Bias, and follows the tendency of chlorophyll-a to occur as pointed out by the in situ data. However, it is possible to observe the performance of the algorithm at different collection points along the river and realise that it can estimate the variation of chlorophyll-a concentration broadly and not more punctual.

Mapping of chlorophyll-a concentrations
The mapping of chlorophyll-a concentrations was performed using the QGIS Desktop 3.4.5; when a clipping was generated referring to the study area, in the images resulting from each calculated equation it was possible to spatially verify the occurrence and distribution of chlorophyll-a along the Poti River in the mentioned biennium.
Maps derived from Sentinel-2A-L2A show details of the spatial variation of chlorophylla concentration, allowing easy identification of areas with high or no chlorophyll-a concentrations, and show that there is a common pattern between estimated and in situ data. At the same time, regardless of the algorithm used, the maps always show the stretch from Point PT-04 to the mouth of the Poti River, approximately 16 km long, as the area with a high concentrations of chlorophyll-a.   Figure 3 shows the change in chlorophyll-a concentration at the beginning of the dry seasons of 2016 and 2017, calculated with data from the Chl-2 algorithm (Page et al., 2018), with emphasis on the section of Point PT-04 to the mouth of the Poti River in the months of

CONCLUSIONS
This study is the first attempt to evaluate the use of Sentinel-2 in Poti River remote sensing. The results show that Sentinel-2 images have great potential for urban river remote sensing, as we were able to retrieve and map chlorophyll-a concentrations by means of routinely used empirical algorithms.
The results generated by the Chl-1, Ha et al. (2017), Chl-2, Page et al. (2018), and Chl-3, Kuhn et al. (2019) equations show the need for calibration of the models used to simulate the Poti River water components. Considering a moderate correlation between the results of punctual quantitative agreement between the in situ data and the values obtained with each equation, the empirical algorithm Chl-2 obtained a better punctual performance than the algorithms Chl-1 and Chl-3.
The empirical algorithm Chl-2 shows a correlation has been established to identify the spatiotemporal variation of chlorophyll-a concentration along the Poti River broadly and not more punctually. The spatiotemporal distribution of chlorophyll-a in maps from Sentinel-2A images are consistent with the pattern of occurrence determined by the in situ data. Therefore, the MSI sensor proved to be a tool suitable for the detection and monitoring of chlorophyll-a concentration along the Poti River.
In addition, it is recommended that a chlorophyll-a concentration monitoring system be implemented using a calibrated algorithm for the Poti River optical properties and Sentinel-2 data in order to improve the environmental management of the river. To enable the monitoring of water quality in urban rivers, especially in the tropical semiarid region of Brazil is a contribution to the conservation of the environment and the sustainable management of this water resource, as it will permit the verification of current and possible future conditions of eventual environmental, social and economic impacts.