Remote sensing-based algorithms for water quality monitoring in Olushandja Dam, north-Central Namibia

Frequent and continuous water quality monitoring of Olushandja Dam in Namibia is needed to inform timely decision making. This study was carried out from November 2014 to June 2015 with Landsat 8 re ﬂ ectance values and ﬁ eld measured water quality data that were used to develop regression analysis-based retrieval algorithms. Water quality parameters considered included turbidity, total suspended solids (TSS), nitrates, ammonia, total nitrogen (TN), total phosphorus (TP) and total algae counts. Results show that turbidity levels exceeded the recommended limits for raw water for potable water treatment while TN and TP values are within acceptable values. Turbidity, TN, and TP and total algae count showed a medium to strong positive linear relationship between Landsat predicted and measured water quality data while TSS showed a weak linear relationship. The regression coef ﬁ cients between predicted and measured values were: turbidity (R 2 ¼ 0.767); TN (R 2 ¼ 0.798,); TP (R 2 ¼ 0.907); TSS (R 2 ¼ 0.284,) and total algae count (R 2 ¼ 0.851). Prediction algorithms are generally best ﬁ t to derive water quality parameters. Remote sensing is recommended for frequent and continuous monitoring of Olushandja Dam as it has the ability to provide rapid information on the spatio-temporal variability of surface water quality.


GRAPHICAL ABSTRACT INTRODUCTION
Fresh water is a finite resource that is essential for human existence (Maestu ; Bhuyar et al. a Therefore, Integrated Water Resources Management (IWRM) is an essential approach as it emphases on effective management of water resources within the basin.
In the Southern Africa Region, pollution of both surface and ground water is on the increase, particularly from mining, agricultural and industrial activities (SADC ).
Man-made reservoirs such as dams and lakes are threatened by nutrient enrichment and heavy metals and their water quality is continuously degraded as a result (Oberholster & Ashton ; Nhapi ; Lehmann ). For example, the water quality in Von Bach and Swakoppoort Dam in the Central Area of Namibia has been reported to be poor due to treated and at times partially treated wastewater from the City of Windhoek and Okahandja town (Lehmann ; NamWater ).
In the north central regions of Namibia, poor sanitation practices are notably high (IWRMPJVN ), with about 67% of the population having no access to improved sanitation facilities (UN 2010). North-central Namibia is more vulnerable to effects of climate change and variability. Combined effects of environmental degradation, social vulnerability to poverty and a changing climate will compromise water and sanitation provision (Angula & Kaundjua ). Flash floods are further predicted to impact overall sanitation and human health conditions (Kaundjua et al. 2012). Therefore, during high rainfall events, cholera outbreak has been reported due to wash away of water and wastewater aided by poor and inadequate sanitation facilities in the areas (UN ). Shuuya & Hoko () found that the quality of water in the Calueque -Oshakati canal is deteriorating from upstream to downstream due to human activities along the canal such as agriculture, settlements and mining. All these activities tend to affect the quantity and quality of water leading to an increase in the cost of purifying water for human consumption. The IWRMPJVN () noted that lack of data, continuous monitoring and poor data management exist in the Namibian water sector. Water quality monitoring in many southern African countries including Namibia has been noted to be hindered by lack of data, capacity and resources (SADC ). Water quality monitoring in Namibia has been noted to be one of the key issues that requires urgent attention by researchers and water managers.
Few studies that are available for the Kunene River and Cuvelai Basins in the north-central regions of Namibia focus on ad-hoc monitoring. Hambabi () and other studies focused on the quality of water in the canals that bring water into the Olushandja Dam and to the treatment plants (Shuuya & Hoko ) and transport water from the dam (SDP ). These studies mainly employ traditional insitu-based methods of determining water quality.
According to the IWRMPJVN (), determining water quality using traditional techniques is costly and may be one of the major contributors to poor monitoring framework not only in Namibia but also in many developing countries in the world. Ritchie et al. () stated that, traditional methods for assessing and monitoring water quality are expensive and time consuming. In addition, they do not give spatial or temporal view of water quality needed for accurate assessment of water bodies, therefore the need for more robust techniques which have a spatial and a temporal dimension (Kallio ).
Thus, this study explored the applicability of remote sensing in combination with insitu-observed measurements to develop algorithms and prediction of selected water quality parameters in Olushandja Dam. Remote sensing based water quality assessment is an economical way to monitor water quality, since it allows routine monitoring of large areas in a short time and on a repetitive basis (Hellweger et al. ; Somvanshi et al. ). According to Ritchie et al. (), the use of remote sensing in water quality dated back to 1970s. Namibia is one of the countries with the most number of sunshine days or cloud free sky which make it easy for optical remote sensing application in the visible and infrared regions of the electromagnetic spectrum.
Information obtained from this study will help the institutions responsible for management of Olushandja Dam such as NamWater and Namibian community-based water management (CBWM) Kelbert () to carry out continuous monitoring of the quality of water in an economical way for decision making.
However the major constraints would be the lack of reliable retrieval algorithms, cost of satellite data and equipment's for in-situ, on-site and laboratory measurements of water quality parameters (Bauer et al. ).

Field data collection
Sampling points (Figure 2) in the dam were selected systematically to reflect on the impacts of major tributaries' landcover activities on the dam water quality. Parameters were also selected taking into account the relationship between pollution sources and different water quality parameters including their potential impacts on human, aquatic life and environmental health. In addition, the selection considers water quality parameters that are known to be successfully retrieved from satellite data. Turbidity, algae and nutrients are known to have been successfully retrieved from

Application of the algorithms and validation
The developed algorithms were applied to resampled multispectral reflectance image bands. Interpolation was performed on the 6 sampling locations to come up with a spatial variation map of each parameter. The Inverse Distance Weighted interpolation (IDW) (Parida et al. ) technique was used for interpolating water quality data at respective point locations to grid points for comparison with satellite retrieved estimates.
Furthermore, the predicted values for each parameter were extracted from every parameter satellite raster map based on GPS location. The predicted values were validated against field measured data to test whether the satellitebased predictions were able to estimate water quality parameters. This was done by simple regression analysis using field data (for 25 January and 10 February 2015) that were randomly selected from main dataset and not used in the development of the algorithms to reduce errors. To assess detailed spatial distribution of water quality in Olushandja Dam, quantitative attribute maps were produced by applying algorithms to the original Landsat datasets.

Field measured data
The mean field measured water quality parameters data acquired from four sampling campaigns (26 Feb-15 April 2015) were calculated at each of the six points. The field data that correspond to the same dates as the satellite data (Table 1) are eventually used for model development.

Remote sensed data
The spectral band reflectance values at each sampling point on selected four images corresponding to field data (26 February and 15 April 2016) are given in Table 2. From the table, low reflectance values (<0.1) on the water body can be noticed and this could be due to the fact that water is relatively clean and this make it have low reflective spectral radiation (He et al. ).

Developed algorithms
Field measured data (Table 1) were regressed with reflectance values (given in Table 2) using a simple linear regression method and algorithms were formulated and given on Table 3. Looking at the R 2 (coefficient of determination) and F-values (which approaches 1 for good performance), the results of regression analysis can be interpreted as follows; all the derived regression models have good regressive correlation for all parameters (i.e. total algae, turbidity, TN, TP, TSS and NH 3 ). Total algae, turbidity, TN, TP and TSS had R 2 values of 0.999, 0.986, 0.987, 0.980, 0.988, and 0.917 respectively. The R 2 values show that there is a high accuracy in predicting all the water quality parameters for the whole dam.
The R 2 values for all water quality parameters were high and close to 1, which shows that there is a significant relationship between satellite data and insitu-observed measurements.  (Table 4) and spatial distribution (Figure 4(a)-4(e)) of the predicted five water quality parameters are presented.

Spatial variation of predicted water quality parameters
The resultant quantitative maps for the five water quality variables (total algae content, turbidity, and concentration of total nitrogen, total phosphorus and total suspended solids) are presented in Figure 4 The resultant maps demonstrated that total algae contents and concentrations TN, and TSS are highly variable over the dam while turbidity and TP had less spatial variance.
Turbidity is found to be above recommended limits of drinking water standards by NamWater as well as for recreational   NamWater values are found to be within the UNECE () standards for maintaining the aquatic life which <300 mg/L.

Validation of algorithms
The relationship between Landsat predicted values and insitu-observed measurements of water quality parameters at each of the 6-sampling points for 15 April 2015 are presented in Figure 5  data. It was found that TP has higher R 2 followed by total algae counts, TN and turbidity respectively. TSS had a low correlation coefficient.

Methodological framework for monitoring water quality
Based on the result of this study, Landsat data has proven its ability to derive water quality parameters, data that are comparable to the insitu-observed measurements. Therefore, a proposed a framework for predicting water qualities for Olushandja Dam is presented in

DISCUSSION Summary
This research found that regression analysis-based retrieval algorithms are ideal for water quality retrieval in Olushandja Dam, North-Central Namibia. Obtaining the spatial variation of water quality parameters allows decision makers to manage the critical resource in near real time. Remote sensing is therefore recommended for frequent and continuous monitoring of Olushandja Dam as it has the ability to provide the information on the spatio-temporal variability of surface water quality. A detailed discussion and future applications of the present work are described in next sections.
Regression analysis-based algorithms for water quality retrieval Insitu based water quality retrieval (Bhuyar et al. a, b)  In addition to the 6 sampling points, the Landsat predicted data at specific sampling points were also compared to water quality standards for the different uses. Turbidity was found to be above the NamWater () of 5 NTU but below the Canadian (2012) (2007)

CONCLUSIONS AND RECOMMENDATIONS
The capacity to predict water quality status under data scarce environments can potentially contribute towards improved water quality management. The study demonstrated that a linear regression statistical approach can be used to develop algorithms for retrieving water quality data from satellite imagery as the correlation between the water quality values retrieved from remote sensing techniques had a high correction with field measured data (R 2 > 0.98). The algorithms developed have a good prediction accuracy for turbidity, total nitrogen, total phosphorous and total algae count as the linear relationship between Landsat predicted and measured values had R 2 of 0.77-0.91 while prediction for total suspended solids showed a weak linear relationship (R 2 ¼ 0.28). The developed of algorithms can be used to predict water quality variables for the whole Olushandja Dam but require more sampling points to further improve the accuracy of regression models. We recommend further development of a GIS web-based application for rapid retrieval and communication of the water quality information by water managers, water users and stakeholders in the Cuvelai Basin in north-central regions of Namibia.