Deep learning-based retrieval of cyanobacteria pigment in inland water for in-situ and airborne hyperspectral data. (March 2020)

• Record Type: Journal Article
• Authors: Yim, Inhyeok; Shin, Jihoon; Lee, Hyuk; Park, Sanghyun; Nam, Gibeom; Kang, Taegu; Cho, Kyung Hwa; Cha, YoonKyung
• Is Part Of: Ecological indicators. Volume 110(2020)
• Publication Date: 2020-03
• Subjects: Cyanobacteria -- Phycocyanin -- Hyperspectral imaging -- Deep learning -- Deep neural networks -- Stacked autoencoder; Environmental monitoring -- Periodicals; Environmental management -- Periodicals; Environmental impact analysis -- Periodicals; Environmental risk assessment -- Periodicals; Sustainable development -- Periodicals; 333.71405
• DOI: 10.1016/j.ecolind.2019.105879
• Languages: English
• ISSNs: 1470-160X

Abstract

Highlights: SAE-DNN is proposed for estimation of phycocianin concentration in inland waters. SAE-DNN outperforms DNN and band-ratio algorithms. SAE-DNNPeaks, a simplified model using selected bands, provides comparable accuracy. Phycocianin in airborne data estimated with SAE-DNNPeaks, with moderate accuracy. Abstract: Worldwide proliferation of cyanobacteria blooms in inland waters not only affects the intended use of water but potentially threatens human and animal health. In this study, a stacked autoencoder-deep neural network (SAE-DNN) was developed to estimate phycocyanin (PC) concentration by using in situ reflectance spectra in productive inland water. The estimated PC using the SAE-DNN was in close agreement with the measured PC, with an R 2 of 0.87, root mean square error (RMSE) of 14.45 μg/L, and relative RMSE of 86.42%. The performance of the SAE-DNN was superior to that of the DNN and band-ratio algorithms. An analysis on the deep spectral features extracted using the SAE yielded the most useful spectral bands, namely 538, 596, and 735 nm, for the retrieval of PC. The estimation accuracy of the SAE-DNNPeaks, using only the aforementioned spectral bands as input variables, was comparable to that of the SAE-DNN, demonstrating that the high-level of abstraction using the SAE facilitated the improvement in feature learning. The application of the SAE-DNNPeaks to airborne hyperspectral image data resulted in an acceptable estimation accuracy, despite a bias toward underestimation, potentially arising from uncertainty associated with atmospheric correction, at high PC concentrations. Our results suggest that simple, empirical-based approaches, such as the SAE-DNNPeaks, have the potential to serve as a rapid assessment tool for the abundance and spatial distribution of cyanobacteria.