Proceedings Article | 11 October 2018
KEYWORDS: Machine learning, Earth observing sensors, Landsat, Satellites, Data modeling, Magnesium, Spatial resolution, Algorithm development, Performance modeling, Coastal modeling
Frequency and intensity of the harmful algal blooms (HABs) increased globally since 1970s. The increase in HABs have negatively affected aquatic ecosystem and aquaculture industry. The economic losses were about $ 1 billion in Europe, $ 100 million in USA and $ 121 billion in Korea per year. There were various field monitoring campaigns for ecological and biological researches. However, traditional HABs monitoring has limitations on both spatial and temporal coverage. In these days, multispectral remote sensing methods using satellite sensors have been widely used to monitor HABs in ocean and coastal areas. However, the satellite systems used in ocean and coastal research, such as MODIS, SeaWiFS and etc. have limitations in study on complex coastline, because of their coarse spatial resolution (~ few km). In this research, we conducted two-year intensive monitoring on the South Sea of Korea from 2016 to 2017 at 62 sampling station and used landsat-8 operational land imager (OLI) satellite that has 30m spatial resolution. We used 4 band (band 1 to 4), 4-band ratio (band 1 over band 3 and 4, and band 2 over band 3 and 4) and mixed dataset of 4 band and 4-band ratio. The empirical OC algorithms showed poor performances, under 0.25 of r-squared. The machine learning techniques, i.e., artificial neural network (ANN) and support vector machine (SVM) were applied to enhance performance of estimating chl-a on landsat-8 application. Parameters for developing ANN and SVM model were optimized using a pattern search algorithm in MATLAB toolbox. All dataset were divided into 80 % of training and 20 % of validation data. In the training step, mixed dataset showed the best performance in both ANN and SVM models, whereas 4-band ratio and 4 band dataset in the validation step showed the best performance in ANN and SVM, respectively. The ANN model showed poor performance in low chl-a concentrations but SVM had more accurate performance in low and mid concentrations. Both models under-estimated chl-a in mid to high concentration range. For the mapping results, the ANN model using 4 band dataset showed very low concentration of chl-a in most of research area, whereas SVM showed high concentration of chl-a in coastal area and bay. The result using 4-band ratio dataset showed similar chl-a distribution in ANN and SVM. For mixed dataset results the ANN model estimated over 8 mg m-3 of chl-a at some of coastal, almost zero in near coastal area and over 2 mg m-3 chl-a concentration for off-shore area. In case of SVM, all region showed approximately 2 mg m-3 of chl-a concentration. Landsat-8 OLI was not proper system for OC algorithms. Machine learning techniques were effective tools for enhancing ocean chl-a estimation performance using landsat-8 OLI. Thus, this study showed potential of landsat-8 OLI application to coastal HAB monitoring.
