Estimating Chlorophyll-a and Dissolved Oxygen Based on Landsat 8 Bands Using Support Vector Machine and Recursive Partitioning Tree Regressions

Abstract

In general, water quality mapping is done by interpolation of in situ measurement samples. Often, these parameters change with time. Due to geographic variability and the lack of budget in Nepal, such measurements are done less often. Remote sensors that collect spectral information continually can be very useful in the regular monitoring of water quality parameters. Landsat Operational Land Imager (OLI) bands have been used to estimate water quality parameters. In this work, we model two water quality parameters: chlorophyll-a (Chl-a) and dissolved oxygen (DO) using sequential minimal optimization regression (SMOreg), which implements a support vector machine (SVM) algorithm and recursive partitioning tree (REPTree) regressions. A total of 19 measurements were taken from Phewa Lake, Nepal and various secondary bands were derived from using Landsat 8 Operational Land Imager (OLI) bands. These bands underwent feature selection, and regression models were created based on selected bands and sample data. The results showed satisfactory modelling of water quality parameters using Landsat 8 OLI bands in Phewa Lake. Due to a limited number of data, cross-validation was done with 10 folds. The SVM showed a better result than the REPTree regression. For future studies, the performance can be further evaluated in large lakes with larger sample numbers and other water quality parameters.