主成分提取在遥感FTIR谱图解析中的应用

建立了基于人工神经网络(ANN)的遥感FTIR谱图解析方法.针对人工神经网络(ANN)训练时间过长和模型"过拟合"的问题,采用偏最小二乘法(PLS)和主成分分析法(PCA),对输入ANN的光谱数据进行了主成分提取,使ANN分析时间从30多分钟缩短为10多秒钟;模型传递技术的引入,克服了遥感FTIR谱图分析中反复建模问题.经过优化的方法,实现了用EPA数据建模,对大气中的四组分混合体系--丙酮、苯、三氯甲烷和甲醇的遥感、实时、准确测定,PLS-ANN模型得到的结果最好,对丙酮、苯、三氯甲烷和甲醇的预测误差分别为0.043,0.031,0.034,0.051,保证了遥感FTIR对大气中有毒气体混合物实时、准确、快速监测.

Principal Component Extraction Used for the Interpretation of RS-FTIR Spectra

A method for interpretation of remote sensing FTIR spectra was set up based on ANN model. Considering long training time and over-fitting problem of ANN, two methods, partial least squares (PLS) and principal component analysis (PCA), were utilized to extract principal components of spectra, process time decrease from about 30 minutes to a few seconds. Meanwhile, the idea of calibration transfer was used to overcome the limitation of calibration model in remote sensing FTIR spectra analysis. With the optimization of ANN model, four-component mixtures of acetone, benzene, chloroform and methanol were predicted in a remote sensing and real-time way while the calibration model was built with EPA data. The best performance was yielded with PLS-ANN model, and the root mean square error (RMSE) of acetone, benzene, chloroform and methanol were 0.043, 0.031, 0.034 and 0.051 respectively, which confirm the real-time, correct and quick analysis of remote sensing FTIR in air monitoring.