高光谱技术结合CARS-ELM的油桃品种判别研究

Study on Varieties Discrimination of Nectarine by Hyperspectral Technology Combined with CARS-ELM Algorithm

摘要:

基于高光谱技术研究竞争性自适应重加权算法（competitive adaptive reweighted sampling，CARS）结合极限学习机（Extreme learning machine，ELM）对油桃品种判别的可行性。本文利用高光谱成像技术选取油桃420~1000 nm的高光谱图像数据，经卷积平滑法（Savitzky-Golay smoothing，SG）、附加散射校正算法（Multiplicative Scatter correction，MSC）、基线校正（Baseline）、变量标准化算法（Standard Normalized Varite，SNV）等预处理方法处理原始数据，通过PLSR模型确定Baseline为最佳预处理方法。采用主成分分析法（Principal Component Analysis，PCA）、连续投影算法（Successive Projections Algorithm，SPA）与竞争性自适应重加权算法等提取的特征波长，建立偏最小二乘（Partial Least Square，PLS）和极限学习机鉴别模型进行比较研究。结果显示：基于CARS算法提取的特征波长构建的CARS-ELM和CARS-PLS模型性能最优。CARS-PLS预测集相关系数（RP）和均方根误差(RMSEP)分别为0.942和0.205；CARS-ELM的RP和RMSEP分别为0.931和0.119。说明CARS是一种有效的提取特征波长的方法，且ELM与PLS对模型的预测能力相当，可见利用高光谱图像技术结合CARS-ELM对油桃的品种判别是可行的。

Abstract:

The varieties discrimination feasibility of nectarines by hyperspectra technology combined with competitive adaptive reweighted sampling (CARS) and extreme learning machine(ELM) were discussed. Hyperspectra imaging technology was used to select the hyperspectra images data of three different types of nectarines in the range of 420 to 1000nmin this study.The raw spectra were processed by Savitzky-Golay, Multiplicative Scatter Correction, Baseline, and Standard Normalized Varite. The most appropriate pretreatment method was confirmed by the PLSR model. In order to extract characterized wavelengths, there are three different approaches have been used: principal component analysis (PCA), successive projections algorithm (SPA), and CARS. Furthermore, partial least square(PLS)and extreme learning machine (ELM) on the basis of three different characterized wavelengths were used to build the model for identifying the species of nectarines. The results of tests implicated that the performance of CARS-PLS and CARS-ELM classification model was optimized using optimal characterized wavelengths of CARS algorithm. The PLS and ELM model correlation coefficient (Rp), the root mean squared error (RMSEP) of prediction set were 0.942, 0.205 and 0.931, 0.119, respectively. It indicated that CARS algorithm was one of the effective methods for exacting the characterized wavelengths,and ELM model prediction ability was equivalent with conventional PLS model. It approves that the hyperspectral imaging technology combined with CARS-ELM is feasible to discriminate the nectarine varietiesn.