Comparison of Multiple Linear Regression and Artificial Neural Network Models in Retrieving Water Quality Parameters Using Remotely Sensed Data.

Abstract

Water quality is an essential resource for the survival and well-being of humans and ecosystems and hence the quality of water is also crucial. In order to know the quality of surface water, water quality parameters are measured traditionally by using in-situ measurements. However, in-situ measurements are time-consuming, labor-intensive, and almost impossible to obtain measurements of the whole water body. Therefore, in this study, we compare the capability of Multiple Linear Regression (MLR) and Artificial Neural Network (ANN) in retrieving water quality parameters from remotely sensed data in inland water, case of Lake Victoria, in Tanzania. The models are used commonly for retrieving remote sensing-based water parameters. The performance of MLR and ANN in retrieving Turbidity and TDS is evaluated. Surface reflectance values from Landsat 8 Operational Land Imager (OLI) sensor images and in-situ data are used to find reliable relationships between Turbidity and TDS. The results indicate that ANN performs better than MLR in retrieving Turbidity and TDS. ANN had an accuracy (R2) of 88.73% and 83.36% in retrieving Turbidity and TDS respectively while MLR had an accuracy (R2) of 66.66% and 78.42% in retrieving Turbidity and TDS respectively. Other criteria that were used for comparison include the standard error, root mean square error, and mean absolute error which both indicated ANN had better performance than MLR. The distribution of Turbidity and TDS mapped in Lake Victoria generally shows that the lake has good water quality as described by WHO standards and could be used for human being consumption. Based on the results attained we recommend, the utilization of ANN and Landsat 8 (OLI satellite images for water quality parameter modeling.