水体透射光谱结合主成分分析(PCA )改进化学需氧量(COD)含量估算研究

化学需氧量（Chemical Oxygen Demand，COD）是水体有机污染的一项重要指标，化学需氧量越高，表示水污染程度越严重。 为了解决传统的COD测量方法耗时较长，不利于快速、实时地获取水体中COD的信息等问题。本文提出了基于透射光谱测量结合主成分分析(Principal Component Analysis, PCA)改进水体COD含量估算模型。具体的，采集100组COD水体光谱信息，分别使用3种不同的高光谱数据预处理方法对光谱数据进行预处理，分析不同预处理方法对模型精度的影响，并基于不同的预处理方法分别建立高斯过程回归模型(Gaussian Process Regression, GPR)和BP神经网络模型，分析不同预处理方法对模型精度的影响；并对各模型结合PCA数据降维方法进行模型的改进，通过比较模型的精度选择最优模型进行水体COD含量的检测。结果显示，相比于原始光谱数据建立的GPR模型和BP神经网络模型，数据预处理后的模型精度明显提升；且结合PCA对预处理后的数据进一步降维处理后，模型精度得到了进一步的提升。其中，基于标准正态变量变换特征结合PCA改进BP神经网络模型基于PCA改进的BP神经网络模型R^2高达0.9940，均方根误差RMSE为0.022540。证明了基于PCA改进的BP神经网络数据降维方法对预处理后的光谱数据进行降维处理，有利于去除光谱中的冗余信息，提取特征信息，可以实现高光谱检测方法可以实现COD含量估算模型的优化，从而为传统COD测量方法存在的问题提出了一种新的解决思路。

Estimation of COD content by transmission spectroscopy combined with PCA

COD is an important indicator of organic pollution in water, and the higher the COD, the more serious the degree of water pollution. To solve the traditional method of measuring COD is time-consuming, not conducive to rapid, real-time access to COD information in the water body and other issues. In this paper, an improved model for the estimation of COD content in water bodies on the basis of transmission spectroscopy measurement combined with principal component analysis (PCA) is proposed. Specifically, 100 groups of COD water body spectral information were collected, and three different hyperspectral data preprocessing methods were used to preprocess the spectral data, and Gaussian Process Regression (GPR) and BP neural network models were constructed based on different preprocessing methods to analyze the effects of different preprocessing methods on the accuracy of the models. In order to analyze effects of different preprocessing methods on model accuracy, GPR and BP neural networks have been constructed based on different preprocessing methods. Compared with GPR model and BP neural network model constructed from original spectrum data, it was found that after data pre-processing, there was a significant improvement in model accuracy, and after further dimension reduction of pre-processing data combined with PCA, there was a further improvement in model accuracy. Among them, the R^2 of the improved BP neural network model based on standard normal variable transformed features combined with PCA is as high as 0.9940, and the RMSE is 0.022540. This proves that the dimensionality reduction of the preprocessed spectral data based on the PCA data dimensionality reduction method helps to remove the redundant information in the spectral data and extract the feature information, and optimizes the COD content estimation model, thereby solving the problems of the traditional COD measurement methods. Thus, a new idea for the solution of the problems that exist in the traditional method of COD measurement is proposed.