FEATURE SELECTION OF OPTICAL SATELLITE IMAGES FOR CHLOROPHYLL-A CONCENTRATION ESTIMATION
M. V. Nguyen1,2, H. J. Chu1, C. H. Lin1, and M. J. Lalu3
- 1Department of Geomatics, National Cheng Kung University, Taiwan
- 2Institute of Geography, Vietnam Academy of Science and Technology, Vietnam
- 3Department of Geomatics Engineering, Institut Teknologi Sepuluh Nopember, Indonesia
Keywords: Feature Selection, Water Quality, Chlorophyll-a, Remote Sensing, Optical Satellite Images, Turbid Inland Water
Abstract. Healthy inland freshwater sources, such as lakes, reservoirs, rivers, and streams, play crucial roles in providing numerous benefits to surrounding societies. However, these inland water bodies have been severely polluted by human activities. Therefore, long-term monitoring and real-time measurements of water quality are essential to identify the changes of water quality for unexpected environmental incidents avoidance. The success of satellite-based water quality studies relies on three key components: precise atmospheric correction method, optimization algorithm, and regression model. Previous studies integrated various algorithms and regression models, including (semi-) empirical or (semi-) analytical algorithms, and (non-) linear regression models, to obtain satisfactory results. Nevertheless, the selection of appropriate algorithm is complex and challenging because of the fact that the changes in chemical and physical properties of water can lead to different method determination. To alleviate the aforementioned difficulties, this study proposed a potential integration which comprises an optimization method for efficient water-quality model selection, ordinary least squares regression, and an accurately atmospheric corrected dataset. Prime focus of this study is water-quality model selection which optimizes an objective function that aims to maximize prediction accuracy of regression models. According to the experiments, the performance of the selected water-quality model using proposed procedures, dominated that of the existing algorithms in terms of root-mean-square error (RMSE), the Pearson correlation coefficient (r), and slope of the regressed line (m) between measured and predicted chlorophyll-a.