Estimation and mapping of water quality parameters using satellite images: a case study of Two Rivers Dam, Kenya

The world's population is fast growing and this keeps on pushing more people to settle in fragile ecosystems mainly the shores of lakes and rivers, and coastal regions (Bar-Massada et al. 2014). The increasing anthropogenic activities in these areas are a major threat to most of the world's freshwater resources (Vörösmarty et al. 2010) and could create irreversible negative impacts, especially in a changing climate. Continuous watershed monitoring is a crucial concept that will help realize sustainable water supplies. According to Najafzadeh & Niazmardi (2021), the quality of surface water plays a key role in the sustainability of ecological systems. Measuring water quality parameters (WQPs) is of high importance in the management of surface water resources. Furthermore, the treatment and supply of water to meet the needs of the various end users require an understanding of the real-time quality of water at the source which influences the choice of chemicals and quantities used in the treatment process.

Eldoret is a fast-growing town with an estimated urban population of 475,716 in 2019 and a growth rate of 3.82% per annum. According to the Eldoret Water and Sanitation (ELDOWAS) Company, the demand for water in Eldoret Municipality is estimated at 60,000 m³/day, against production of 36,400 m³/day (Kimutai et al. 2018). According to Kibii et al. (2021), mismanagement in the catchment is partly responsible for the huge disparity between demand and supply due to the recent conversion of forested land into subsistence agriculture. This has led to flash floods, erosion, and sedimentation which decrease the quality of surface water. In addition, competing users and uses have contributed to a substantial increase in freshwater requirements. Furthermore, factors such as climate change, population growth, and inadequate conservation practices in the catchment negatively impact the water quality as exhibited through increased turbidity and algal blooms (Ontumbi et al. 2015; Barasa & Perera 2018).

Monitoring surface water quality is crucial mainly in the context of increasing freshwater demands and wastewater discharged to the environment (Chen & Han 2018). The traditional approach of water quality monitoring (WQM) entails the collection of samples in the field followed by a water quality analysis in the laboratory. However, the approach is time-consuming, labor intensive, and it does not give the spatial-temporal variations of the water quality indices (WQIs). Furthermore, reliance on conventional methods also limits the possibility of monitoring, forecasting, and managing entire water bodies due to the large extent of the water surface, lack of spatial-temporal data on a regional scale, and geographical limitations (Gholizadeh et al. 2016).

According to Najafzadeh et al. (2021), WQIs are crucial in describing the essential characteristics of water pollutants and this creates the need for accurate predictions of WQIs in order to gain insights into the patterns of pollutants in natural streams. Furthermore, Najafzadeh et al. (2021) also note that one of the most difficult issues in the studies of water quality specifically, surface water resources, is getting an accurate estimate of WQIs. Even though there are numerous conventional methodologies for evaluating the WQIs, the limitations that exist among the traditional models have brought the need to employ data-driven models (DDMs) in assessing the WQIs of natural streams. The WQM challenges can also be overcome by using satellite images which avail a smart, rapid, and low-cost WQM tool.

The advancements in computer science and remote sensing techniques have made remote sensing find wider applications in WQM (Usali & Ismail 2010; Gholizadeh et al. 2016). The use of remote sensing techniques and satellite images allows for continuous WQM over larger spatial and temporal scales thus improving the water management practices for vast geographical areas (Japitana & Burce 2019). The concept could also be extended to determine the impacts of anthropogenic activities in different catchments, pollution management, and watershed management. The development of WQM systems that incorporate remote sensing increases the efficiency with which individuals respond to emergency ecological challenges such as point and non-point pollution, algal blooms, and floods. Real-time measurements also enable data to be analyzed rapidly and effectively while limiting errors that come with sample collection and laboratory analysis.

Ouma et al. (2018) note that empirical models leverage bivariate and/or multiple regressions between data acquired from sensors and WQPs measured in situ by correlating the sensor radiance values and their band combinations with WQPs collected and measured based on the sensor overpass schedule. For instance, empirical multivariate regression modeling (EMRM) simulations are done to determine the multivariate correlations between the reflectance from sensor bands and WQPs measured in situ. Furthermore, Najafzadeh et al. (2018) also highlight the need for evolutionary computing-based formulations including the application of equations extracted from gene expressive programming, and evolutionary polynomial regression (EPR) in the prediction of WQPs. This is in line with the increasing recommendation over the last decade to use artificial intelligence models in the prediction of WQPs.

Landsat-8 Operational Land Imager (OLI) is one of the intelligent tools that avails a simple, automated, fast, inexpensive, and noninvasive technology for operational and productive aquatic environmental monitoring (Garaba et al. 2015). Furthermore, the assessment of Landsat-8 imagery also allows for the identification of the optically active water constituents based on their interaction with light and the subsequent energy change of the incident radiation reflected from the water body (Garaba et al. 2015). Numerous algorithms have been developed based on Landsat data for the retrieval of WQP values from remotely sensed imagery. Thus, Landsat-8 OLI is one of the tools that could be used to enable accurate and routine monitoring of water bodies in line with the need for sustainable management of water resources.

According to Ouabo et al. (2020) suitable interpolation techniques can be applied to correctly sampled data in order to make inferences on the distribution and variability of the WQPs at the unsampled locations. However, there is a need to have detailed information on the distribution of WQPs in the reservoir so as to make precise water quality predictions for a specific point in the reservoir. A study by Murphy et al. (2010) compared the performance of ordinary Kriging, inverse distance weighting (IDW), and universal Kriging for spatial interpolation of WQPs. The Kriging-based methods gave better estimates compared to the IDW method with an accuracy of more than 10%.

This study presents a smart WQM approach for the assessment and estimation of WQPs (turbidity, total suspended solids (TSS), and chlorophyll-a (Chl-a)) in Two Rivers Dam, Uasin Gishu County in Kenya. The study used satellite images acquired from Landsat-8 OLI. Ordinary Kriging was applied to estimate the WQPs at the unsampled locations in the reservoir from which the spatial distribution maps were developed.

The data used in this study were collected from the 13 sampling points, L1–L13 (Figure 1). A high concentration of points was near the edges because of the highly variable water quality characteristics at these locations.

The study used Landsat-8 OLI images acquired between 25 November 2020 and 28 January 2021. The selection of in situ sampling dates coincided with the satellite overpass schedule to a tolerance of ±1 day.

Image processing was done using the Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) model where the digital numbers were converted to top-of-atmosphere (TOA) radiance through radiometric calibration. The TOA radiance was then converted to TOA reflectance through atmospheric correction and the surface reflectance was obtained after Dark Object Subtraction (DOS) (Ouma et al. 2018). After the DOS, the region of interest (ROI) for each of the images was then extracted and used in the subsequent processing steps. The Two Rivers Dam shapefile was then used to extract the ROI from the processed images and the ROI used for the subsequent processing.

A total of 78 water samples were collected for the entire sampling duration and the standard laboratory protocols were applied in determining the concentration of turbidity, TSS, and Chl-a. For each of the three sampling dates, two water samples were collected at each of the 13 points. Each sample and its replicate were then tested for turbidity, TSS, and Chl-a, and the average value was recorded. Turbidity measurements were done in the laboratory using the Hanna Portable turbidity meter (model HI98703) and TSS was determined by the gravimetric method (APHA 1975). Chl-a was determined by the spectrophotometric method where the optical density of the extracted Chl-a was measured at four wavelengths (750, 663, 645, and 630 nm) and the resulting concentration was determined based on the SCOR-UNESCO's equations (SCOR-UNESCO 1966). A GPS receiver was used to locate the sampling points during sample collection, thus, enabling meaningful seasonal inferences to be made for the specific locations.

The correlation analysis entailed overlaying the sampling points on the ROI extracted from the processed satellite images for each of the sample collection days. An average spectral reflectance of 3 × 3 pixel neighborhood configuration as proposed by Reddy (1997) was used in order to reduce errors in locating the sampling sites, correlate the reflectance and the WQPs as well as address the high water quality variability since the sampling points were close to the edges of the reservoir. Thus, the 3 × 3 window could include the shallow water near the banks. To convert the surface reflectance values to remote sensing reflectance (Rrs), the surface reflectance values were divided by π (Moses et al. 2015).

The EMRM approach was used to correlate remote sensing reflectance and the WQPs measured in situ as described in Ouma et al. (2018). The band combinations considered for EWRM analysis were single band, band ratio, linear, and mixed combination.

The predicted and laboratory-measured water quality values were compared based on the EWRM algorithm and the equations with the highest R² values were selected (Wang et al. 2006). Eight sampling points were used in the development of the model. The appropriate regression equation was selected based on the value of the coefficient of correlation (R) which was used as a measure of accuracy for the derived equations and five points were used for model validation. The accuracy of the regression results was then determined using the coefficient of determination, Pearson correlation coefficient, mean absolute error or bias, and normalized root mean square error (NRMSE) estimators.

The average in situ turbidity for the entire sampling period varied between 4 and 17 NTU with an average of 7.69 NTU. The reservoir turbidity was generally low since sampling was done during the dry season. The sampling was done between November and January which is a dry period meaning there was no sediment inflow from rainwater discharge into the reservoir. Sediment loads could have also been reduced as a result of plain sedimentation which refers to the quiescent settling of water in a reservoir for extended durations without the aid of chemicals especially when the water source is polluted or highly turbid (Mehdinejad et al. 2012). The concept is more like natural water treatment that results in the settlement of suspended solids, removal of color, hardness reduction, breakdown of organic chemicals, and unfavorable conditions that lead to the death of pathogens. The reflectance from the blue, green, and red bands yielded the highest correlation coefficient between in situ and Landsat-derived water quality values as shown in Table 1. Ouma et al. (2020) also demonstrated that turbidity could be estimated using remote sensing by utilizing the green, blue, and red bands of Landsat-8 OLI. Similar results were also obtained by Lotfi et al. (2019) with the highest correlation obtained between the reflectance values of red and blue bands and in situ turbidity. The significance of the red and blue band in the estimation of turbidity is also emphasized in a study by Kalele (2019) where the best-performing model was a combination of the reflectance values of the red and blue bands with R and RMSE values of 0.841 and 0.828, respectively. Furthermore, the model validation dataset resulted in R and RMSE values of 0.832 and 0.430, respectively. The results from this study show that turbidity for inland waters can be estimated using remote sensing reflectance values from the visible bands of the satellite imagery.

The average in situ TSS for the entire sampling period varied between 247 and 321 mg/L with an average of 277.91 mg/L. The highest concentration of TSS was recorded at the points where River Endoroto and River Ellegerini entered the reservoir since the inflow from the two rivers agitated and suspended the settled sediments from the bottom of the reservoir. Based on the EWRM algorithmic approach shown in Table 1, TSS was best estimated from the coastal aerosol, blue, green, and red bands.

The results in Table 1 relating B3 (green) and B2 (blue) gave an R² value of 0.788. This can be compared with the results of Jaelani et al. (2016) where the logarithmic regression algorithm based on the band ratio of the remote sensing reflectance of B2 (blue) to B3 (green) gave an R² value of 0.79. The concept was also established in a study by Ouma et al. (2020) where the liner regression model from the band ratio between the B3 (green) and B2 (blue) resulted in an R² value of 0.9249. From this study, the single B4 (red) regression model yielded the highest R² value of 0.853. This can be compared to the study by Yanti et al. (2016) where it was established that the estimation and mapping of TSS concentration can be done using a single B4 (red) band. In the study, the single band linear regression model relating in situ TSS to remote sensing reflectance (Rrs) of B4 (red) gave an R² value of 0.5431. In comparison with the other studies, Yanti et al. (2016) concluded that the red band alone is not that informative in the retrieval of TSS. However, in this study, the red band alone was quite informative in the retrieval of TSS from the reservoir. However, just like Yanti et al. (2016) suggested, combining the red band and other visible bands proved to be quite effective in the estimation and mapping of the WQPs in this reservoir. The results from this and similar studies show that TSS can be estimated with relatively high accuracy from the visible bands of satellite imagery. Generally, the nRMSE and Bias errors for this study were also lower compared to the TSS values measured in situ and this proves that the method is sufficient for estimating the TSS concentration of inland waters.

During the entire sampling period, the average in situ Chl-a ranged between 23.58 and 83.15 mg/L with an average of 46.51 mg/L. From Table 1, the highest concentration of Chl-a was recorded in the same regions where high values of TSS were observed. The inflow of water from River Endoroto and River Ellegerini increases the concentration of particulate matter and nutrients at these points. This is because farming is the leading economic activity around the reservoir and this means that the observed Chl-a concentrations can be linked to diffuse pollution by fertilizer leachate from the nearby farms, specifically, an influx of total phosphorous and total nitrogen which are the main variables that contribute to nutrient enrichment. Consequently, the concentration of Chl-a which is the response variable increases. The problem is further worsened by the rainy season which facilitates significant nutrient runoff followed by a dry season, which provides perfect conditions for algae incubation (KDHE 2011). Furthermore, the increased concentration of particulate matter provides attachment sites for the algae and this enables the algal bloom concentration to be propagated thus leading to the observed high concentration of both TSS and Chl-a in the same regions.

The Rrs based on coastal aerosol, green, and red bands gave the best estimate of Chl-a. Similar results were also obtained by Jaelani et al. (2016) where Chl-a concentration retrieval algorithms based on band ratios involving B1, B2, B3, and B4 gave a high determination coefficient (R² > 0.5). The study by Watanabe et al. (2015) where the Rrs spectra computed from in situ data showed high absorption at the blue (B2) and red (B4) spectral regions with the reflection peak being at the green (B3) region and the beginning of the near-infrared region close to the end of the red edge (B4) also confirms that B2, B3, and B4 can be used to estimate Chl-a from Landsat OLI images. As opposed to the study by Lai et al. (2021) which states that the best band combination for the retrieval of Chl-a is that which includes the blue and near-infrared bands, the near-infrared band is not that informative in the retrieval of Chl-a in this reservoir. Regardless, Lai et al. (2021) also acknowledge that if only the near-infrared and blue bands are used for Chl-a retrieval, then the correlation is not ideal. Like most of the cited studies, the current study showed that satellite data such as that from Landsat-8 OLI coastal aerosol, blue, green, and red can be used for estimating and monitoring the seasonal variations of Chl-a in reservoirs and this could help detect, in advance, the occurrence of possible algal blooms.

The validation of the developed regression algorithms was done using data from five sampling stations (L1, L2, L4, L5, and L9 in Figure 1). Table 2 presents calibration and validation results.

The EMRM algorithm used for the prediction of WQPs is a DDM and DDMs have been frequently used to assess the water quality index (WQI) for natural streams (Najafzadeh et al. 2021). The results from all the sampling dates show that TSS had the highest variation in concentration followed by Chl-a and turbidity. Based on the standard deviation (SD), coefficient of variance (CV), and standard error (SE) metrics, the satellite imagery tended to mostly underestimate the concentration of the WQPs but with a very small margin as seen through the low coefficient of variation values. The accuracies obtained based on the EMRM can be compared to the results obtained by Najafzadeh et al. (2021) who used four well-known DDMs including EPR, M5 Model Tree (MT), Gene-Expression Programming (GEP), and Multivariate Adaptive Regression Spline (MARS) for the prediction of the WQI in Karun River, Iran. The number of DDMs feeding-input variables was controlled through techniques like Forward Selection (FS), and Gamma Test and the FS-M5 MT gave the best estimate of the WQI. Even though in this study the number of input variables was not controlled because of their low numbers (only three inputs specifically turbidity, TSS, and Chl-a), great accuracies were still achieved as shown through the validation results in Table 2. This shows that the EMRM approach is a reliable DDM for the estimation of WQPs from Landsat-8 OLI. Najafzadeh & Niazmardi (2021) also developed a Multiple Kernel-Support Vector Regression (MKSVR) algorithm to estimate chemical oxygen demand (COD) and biological oxygen demand (BOD) of Karun River, Iran using different WQPs as input variables. The MKSVR model performed best in estimating BOD with a correlation coefficient R of 0.8 and RMSE of 4.76 mg/l. In comparison with this study, the MKSVR model is more accurate since the EWRM algorithm gave lower correlation coefficient values as shown in Table 1. Even so, the relatively high accuracies could also be attributed to the high number of input parameters considered all of which affect water quality. The idea of using more input parameters to increase the accuracy of WQPs predictions is also supported by the findings from the study by Najafzadeh et al. (2018) where nine input parameters (specifically, Ca²⁺, Na⁺, Mg²⁺, NO²⁻, ⁠, , EC, PH, and turbidity) were used to estimate BOD, COD, and dissolved oxygen using evolutionary computing-based formulations. All three models tested achieved high-performance accuracies as indicated by correlation coefficient (R) of 0.86, 0.76, and 0.84 for GEP, MT, and EPR, respectively.

In Figure 3, the actual and predicted values followed a similar trend line with most points coinciding with actual and predicted turbidity and TSS values. For Chl-a, notable variations were at points (4, 5, and 12) and points (6 and 7) where Landsat-8 values were overestimated and underestimated, respectively, by a significant margin. Figure 4 shows that there was a significant variation of the actual and predicted values specifically for turbidity and Chl-a. Landsat underestimated the turbidity values while Chl-a values were overestimated. However, there was a slight variation between Landsat-predicted and in situ TSS since Landsat estimated the TSS values with a coefficient of variation of less than 10%.

The model results can be compared with the results from the study by Nafsin & Li (2022) which investigated the effectiveness of four stand-alone machine learning (ML) algorithms and six novel hybrid algorithms in predicting the 5-day BOD of Buriganga River, Bangladesh. The Random Forest-Support Vector Machine (RF-SVM), Artificial Neural Network-Support Vector Machine (ANN-SVM), and Gradient Boosting Machine-Support Vector Machine (GBM-SVM) achieved high prediction accuracies of 91, 89.6, and 88.8%, respectively. This means that the ML algorithms, just like the EWRM algorithm, can also be used to improve the accuracy of water quality parameter predictions from satellite imagery. The high prediction accuracies could significantly reduce the coefficient of variation between in situ and Landsat-predicted WQPs.

Overall, Landsat-8 OLI performed well in the prediction of WQPs with reasonable variation based on the SD, CV, and SE metrics. Generally, the effectiveness of atmospheric correction plays a great role in dictating the accuracy of water quality modeling (Bonansea et al. 2019). It is clear that satellite images like those obtained from Landsat-8 OLI can be used as a cost-effective and high-frequency tool for monitoring the water quality of inland waters if adequate radiometric and atmospheric corrections are done.

It is important to determine the spatial distribution of the WQPs in order to visualize the variation in water quality for the entire reservoir from the sampled locations. The spatial maps for the distribution and variability of the observed and estimated WQPs were developed using ordinary Kriging to enable further model performance analysis.

The algorithmic models developed in this study could be used for the prediction and mapping of turbidity, Chl-a, and TSS in Two Rivers Dam reservoir.

The study evaluated the performance of Landsat-8 OLI in predicting the turbidity, TSS, and Chl-a for an inland water reservoir based on in situ measurements at specified sampling points in the reservoir. The results revealed that the mean values of laboratory-measured turbidity, TSS, and Chl-a were 7.69 NTU, 277.9 mg/L, and 46.51 mg/L, respectively and this was highly comparable to Landsat-8 estimated values of 7.22 NTU, 277.71 mg/L, and 46.51 mg/L, respectively. For all the samples, turbidity was estimated using a polynomial regression model with both R² and Pearson correlation coefficient (r) greater than 75%. TSS was best estimated by exponential and polynomial regression models with respective mean R² and r of 0.809 and 0.721. Chl-a was best estimated using polynomial regression models with mean R² and r of 0.803 and 0.731, respectively.

From the study, it has been shown that satellite images including Landsat-8 OLI images avail a cheap and cost-effective tool for reservoir WQM and management. The remote sensing approach ensures continuous water quality assessment and/or management and increases spatial-temporal reservoir monitoring. However, in order to improve the effectiveness and reliability of Landsat-8 OLI in water quality parameter retrieval, it is recommended that the development of the model coefficients be based on data collected on a larger extent of the reservoir for the different seasons in a year. In line with this, model transfer functions should also be developed to enable the algorithms to be used for water quality predictions in other reservoirs within the same location. To apply these models to other localities, the model coefficients must be revised in line with the reservoirs’ hydrological characteristics and the seasonal variations in the specific climatic and hydrological conditions.

The work reported here was undertaken as part of the Building Capacity in Water Engineering for Addressing Sustainable Development Goals in East Africa (CAWESDEA) project which is part of the IDRC funded program on Strengthening Engineering Ecosystems in sub-Saharan Africa. CAWESDEA Project is led by Global Water Partnership Tanzania in collaboration with Makerere University (Uganda), Moi University (Kenya) and University of Dar es Salaam (Tanzania). We acknowledge the support from the Eldoret Water and Sanitation Company (ELDOWAS) for hosting the research reported herein.

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

The authors declare there is no conflict.