均值+方差二维表征高光谱信息的苹果腐败预警方法

为实现苹果在贮藏过程中有效的腐败预警，提出一种基于高光谱图像灰度值均值融合方差的二维高光谱信息表征方法，并构建了苹果样本的巴氏距离（BD）预警模型。首先，为获得有效的光谱信息，对高光谱图像的感兴趣区域（ROI）进行了选择；同时，为减少噪声影响，通过对比分析6种光谱信息的预处理方法，最终采用Savitzky-Golary(SG)平滑方法，分别对均值和方差表征的两种光谱信息进行全波段（371.05～1 023.82 nm）光谱曲线降噪处理。其次，为获得特征波长，从降噪后的光谱曲线中运用连续投影算法（SPA）结合样本色调角和失水率2个理化指标，提取了高光谱图像共同的特征波长，分别得到了两种表征方式下的7个（均值表征）和8个（方差表征）特征波长。然后，通过分析样本色调角随贮藏天数变化的折线图，确定了图中发生明显转折的数据点所对应的贮藏日期，并结合样本贮藏期间实际观察的情况，初步界定第21贮藏日为样本腐败的基准日。另外，依据苹果样本表皮叶绿素特征吸收波长（675 nm左右），绘制出平均光谱反射值变化趋势图，发现趋势图在第21日升至最高点，吻合色调角的分析结果，这表明样本确实从第21日开始腐败。因此，第21贮藏日对应的特征波长可用来建立腐败基准日的光谱信息表征向量。最后，分别建立基于均值表征、方差表征及二者相融合表征下的光谱信息巴氏距离腐败预警模型。结果表明：基于均值融合方差的光谱表征信息所建立的预警模型相较于它们各自建立的预警模型，波动性进一步减弱，可更好地反映苹果样本在贮藏过程中接近腐败的程度。因此，融合灰度值均值和方差的光谱表征信息更能全面的表征苹果贮藏过程中的品质变化，模型预警的稳健性及泛化能力更强，为利用高光谱图像信息对苹果贮藏进行腐败预警提供了新思路。

Early Warning Method of Apple Spoilage Based on 2D Hyperspectral Information Representation With Pixel Mean and Variance

To effectively realize the early warning of apple spoilage during storage, a 2D hyperspectral information representation method based on the mean fusion variance of the hyperspectral image pixel grey value is proposed, and the early warning model of apple samples based on Bhattacharyya distance (BD) is constructed. Firstly, to obtain effective spectral information, the hyperspectral image’s region of interest (ROI) was selected. At the same time, through the comparative analysis of 6 kinds of original spectrum preprocessing methods, and the full-band (371.05～1 023.82 nm) spectral curves represented by the pixel mean and variance were smoothed Savitzky-Golary (SG) for noise reduction, respectively. Secondly, the successive projection algorithm (SPA) combined with the two physical and chemical indexes of sample hue angle and water loss rate was used to extract the feature wavelengths spectrum information, and 7 (pixel mean representation) and 8 (pixel variance representation) common feature wavelengths in the two representation methods were extracted. Then, by analyzing the change of the sample hue angle with the storage time, the storage data corresponding to the data point with a significant turning point was determined and combined with the actual observation during the storage period of the sample, the 21st storage day was preliminarily defined as the spoilage benchmark of apple samples. In addition, according to the characteristic absorption wavelength of the chlorophyll of the apple samples (675 nm or so), the average spectral reflectance change trend graph was drawn, and it was found that the changing trend rose to the highest point on the 21st day, which was consistent with the hue angle analysis result. It shows that the apple samples were indeed spoilt from the 21st day. Thus the spectral information of the 21st storage day corresponding to feature wavelengths can be used as the spectral feature vector of the spoilage benchmark day. Finally, the early warning models of Bhattacharyya distance spoilage based on the mean pixel representation, variance representation and the fusion of the two representation variables were established, respectively. The results show that the early warning models based on the spectral representation information of the pixel mean fusion variance have further reduced volatility compared with their respective early warning models and can better reflect the degree of spoilage of the apple samples during storage. Therefore, the spectral feature information fused with the mean and variance of pixel grey value can more comprehensively characterize the quality changes of apples during storage, and the robustness and generalization ability of the early warning model is strong. The research results provide a new idea for using hyperspectral image information to early warning apple storage spoilage.