Hyperspectral Remote Sensing of the Pigment C-Phycocyanin in Turbid Inland Waters, Based on Optical Classification

Abstract

Pigment C-phycocyanin (C-PC) is a useful indicator for the presence of cyanobacteria in inland waters, which has been well known as a phytoplankton group with many negative effects on human, animal, and aquatic ecosystem health. In recent years, the remote detection of the C-PC concentrations for inland waters has received much attention. However, their accurate quantification by means of remote sensing is still a challenge due to the significant bio-optical complexity of turbid inland waters. In this paper, three typical turbid inland lakes in China were investigated through in situ observed data sets containing optical and water quality parameters. By using a recently proposed <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TD</i> 680 optical classification method, all collected samples were first classified into three types. For each type of water, we determined specific spectral sensitive regions for the pigment C-PC. Then, we developed three type-specific support vector regression (SVR) algorithms and an aggregated SVR algorithm. The performances of these algorithms were evaluated through the validation data sets. The results show that the type-specific algorithms generally have significantly improved performance over the aggregated SVR algorithm. Their assessment errors [mean absolute percentage error ( <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAPE</i> ) and root-mean-square error ( <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rmse</i> )] were as follows: 1) <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAPE</i> = 15.6% and <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rmse</i> = 30.6 mg·m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> for Type 1 water; 2) <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAPE</i> = 47.1% and <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rmse</i> = 61.5 mg·m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> for Type 2 water; and 3) <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MAPE</i> = 26.4% and <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rmse</i> = 19.1 mg·m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> for Type 3 water. The findings in this paper demonstrate that a prior water classification is needed for the development of accurate C-PC retrieval algorithms. This paper provides a valid strategy for improving C-PC estimation accuracy and enhancing algorithm commonality for optically complex turbid waters.