Using multispectral remote sensing for mapping and monitoring water quality at the Vaal dam

Abstract

This project aimed to test the use of two multispectral sensors in estimating and mapping chlorophyll-a and turbidity concentrations, at the Vaal Dam, South Africa. Landsat 8 and Sentinel-2 satellite data were acquired on the 04/10/2016 and 13/11/2016 respectively. Image processing was carried out using atmospheric correction by applying the FLAASH model on the calibrated radiance to obtain atmospherically corrected Landsat 8 images. With Sentinel-2 data, the atmospheric correction was performed using Sen2cor from the Sentinel toolbox to obtain a geometrically corrected Sentinel-2 multispectral image. Band reflectance values were extracted from the two remotely sensed data, and laboratory measurements for chlorophyll-a and turbidity concentrations were obtained from samples collected from 23 sampling points at the dam on the 26-28 October 2016. The remotely sensed data were then cross-validated with the field data in mapping and predicting chlorophyll-a and turbidity concentrations. Two regression models were used in this study; multiple stepwise linear regression, and random forest regression models, to predict chlorophyll-a and turbidity concentrations from the remotely sensed data. The performance and accuracy of these regression models were evaluated by correlating the predicted against the observed chlorophyll-a and turbidity concentrations from the satellite data. Random forest regression gave a higher performance and accuracy than stepwise multiple regression based on their R squared and RMSE values. In mapping chlorophyll-a and turbidity concentrations from both remotely sensed datasets, stepwise linear regression analysis was used to derive estimates for chlorophyll-a and turbidity concentrations. The high estimate values were multiplied with their corresponding bands and added up with intercepts (the expected mean values for chlorophyll-a and turbidity) from coefficients derived from the regression analysis. An equation was then developed using the raster calculator in ArcGIS to map chlorophyll-a and turbidity concentrations from the remotely sensed data for the entire dam. The test of the new generation multispectral sensors Landsat 8 and Sentinel-2 using random forest and multiple stepwise regression models in mapping and predicting chlorophyll-a and turbidity concentrations at the dam was a success. The random forest model gave a better performance than multi stepwise regression. In some cases, the performance of the models was poor as a result of the field data collection date not coinciding with the date of satellite data collection