인공신경망 알고리듬을 이용한 Cochlodinium polykrikoides 적조 탐지

Abstract

This study is to develop detection algorithms for the blooms of Cochlodinium polykrikoides, which has caused serious HABs for more than two decades in Korean waters. Various algorithms have been proposed to detect HABs. However, most of those algorithms are empirical and therefore limited in their applicability in the coastal areas where bio-optical conditions are highly variable spatially and temporally. In a previous study[1], we have shown that remote sensing reflectance of Cochlodinium polykrikoides exhibit a distinctive depression in the blue-green wavelength band based on a large data set of remote sensing reflectance simulated using HydroLight and IOCCG data. We also showed that Cochlodinium polykrikoides can be clearly separated from unspecified phytoplankton assemblages in the two waveband ratio space (Rrs(555)/Rrs(531), Rrs(488)/Rrs(443)). Based on this, we are developing neural network algorithms for in-water and satellite applications. Our preliminary test using simulated and satellite data shows that the success rate of the in-water algorithm ranges from 71.6% (chl-a > 5 mg m-3) to 89.8% (chl-a > 30 mg m-3). We also compare the performance of neural network algorithms for satellite data.hose algorithms are empirical and therefore limited in their applicability in the coastal areas where bio-optical conditions are highly variable spatially and temporally. In a previous study[1], we have shown that remote sensing reflectance of Cochlodinium polykrikoides exhibit a distinctive depression in the blue-green wavelength band based on a large data set of remote sensing reflectance simulated using HydroLight and IOCCG data. We also showed that Cochlodinium polykrikoides can be clearly separated from unspecified phytoplankton assemblages in the two waveband ratio space (Rrs(555)/Rrs(531), Rrs(488)/Rrs(443)). Based on this, we are developing neural network algorithms for in-water and satellite applications. Our preliminary test using simulated and satellite data shows that the success rate of the in-water algorithm ranges from 71.6% (chl-a > 5 mg m-3) to 89.8% (chl-a > 30 mg m-3). We also compare the performance of neural network algorithms for satellite data.