In order to identify the moldy tea leaves in a fast and nondestructive way, a method involving wavelet coupled with k-nearest neighbor (WKNN) was proposed to select effective characteristic wavelengths in this paper. The hyperspectral imaging of 300 dried tea samples with 3 different mildew degrees (contrast check, mild moldy and severe moldy) were obtained using hyperspectral data acquisition device. Besides, food microbiological examination results showed that mold count and total numbers of colony increased with the increase of storage time, temperature and humidity. Roughness penalty smoothing (RPS) algorithm was used to preprocess the raw spectra. Afterwards, WKNN was applied to select the optimal wavelengths of spectral data by using db4, db6, sym5, sym7 as wavelet basis functions, respectively. In addition, five layers of wavelet decomposition were adopted based on different wavelet basis functions. Linear discriminant analysis (LDA) algorithm was used to build the classification models based on preprocessed spectra feature in characteristic wavelengths. The results showed that four optimal prediction models were optimal decomposition level in each wavelet basis function. In addition, the best performance model among all LDA models achieved an identification rate of 100% in the calibration set and 98.33% in the prediction set, in which db4 was used as wavelet basis function and the optimal wavelet decomposition level was 2. WKNN algorithm can effectively achieve the best wavelet decomposition layer and the best wavelengths. WKNN algorithm combined with NIR hyperspectral imaging technology can realize the effective wavelength extraction and classification of dried tea with different mildew degrees.

为了快速，无损地识别发霉的茶叶，提出了一种基于小波和k近邻（WKNN）耦合的方法来选择有效的特征波长。使用高光谱数据采集设备获得了300种具有3种不同霉变程度（对比度检查，轻度发霉和重度发霉）的干茶样品的高光谱成像。此外，食品微生物学检查结果表明，霉菌数量和菌落总数随贮藏时间，温度和湿度的增加而增加。粗糙度惩罚平滑（RPS）算法用于预处理原始光谱。然后，分别使用db4，db6，sym5，sym7作为小波基函数，使用WKNN选择光谱数据的最佳波长。此外，基于不同的小波基函数，采用了五层小波分解。使用线性判别分析（LDA）算法基于特征波长中的预处理光谱特征建立分类模型。结果表明，在每个小波基函数中，四个最优预测模型是最优分解水平。此外，在所有LDA模型中，最佳性能模型在校准集中的识别率达到100％，在预测集中的识别率达到98.33％，其中db4作为小波基函数，最优小波分解级别为2。WKNN算法可以有效地获得最佳的小波分解层和最佳的波长。