简要介绍了偏最小二乘法的原理、算法及优点。将该方法应用于黄海和南海二类水体光谱的水色要素反演,交叉检验结果表明反演精度高,预报相对误差不超过38％。该方法应用于加有5％随机噪声的人工合成光谱的水色要素反演,结果表明模型的稳健性强,预报相对误差不超过5％。研究结果表明,偏最小二乘法适合于处理变量多样本数又少的问题,适合于从二类水体光谱中提取水色要素信息。

It is generally recognized that Case 2 waters are more complex than Case 1 waters in their composition and optical properties. The standard algorithms (usually band ratio) in use today for chlorophyll retrieval from spectral data break down in Case 2 waters. Hyperspectral ocean color sensing may be necessary for Case 2 waters' constituents retrieval. However, hyperspectral data are usually highly correlated and statistical algorithms such as principal component inversion have been employed in ocean color sensing. In the present paper the principle, algorithm and advantage of another statistical algorithm-partial least squares regression (PLS) are briefly described. Then PLS is applied to the retrieval of oceanic color constituents from China Yellow Sea and South China Sea field reflectances, which are typical of Case 2 waters. Cross-validation of PLS analysis shows that the retrieval accuracy is good and the predicted relative error of chlorophyll-a is less than 37% . In order to check the robusticity of the PLS inversion model, PLS is also applied to the retrieval of oceanic color constituents from computed reflectances to which 5% noise is added randomly. The cross-validation results of PLS analysis on simulated data show that the model is robust and the predicted relative error of the three components (chlorophyll-a, Total Suspended Matter and Yellow Substance) is less than 5%. Pre-processing of data is essential for the constituents' concentration ranging over several magnitudes. As an empirical algorithm, the training data set for PLS should be typical that the data points distribute uniformly in the concentration range. It is suggested that PLS be suitable for the regression problems which have a few observations but a lot of spectra variables, e. g. the retrieval of oceanic color constituents from Case 2 water reflectance.