利用可见/近红外光谱透射技术检测温州蜜柑含水率。采用微分处理（differential processing,SD）、多元散射校正（multivariate scattering correction,MSC）、标准正态变换（standard normal variate,SNV）、SG卷积平滑以及标准化等预处理方法比较建立的偏最小二乘回归模型（partial least squares regression,PLS）的拟合准确度，并确定最佳预处理方法，同时采用竞争性自适应重加权采样算法（competitive adaptive reweighted sampling algorithm,CARS）提取特征波长，以此建立基于柑橘含水率的PLS模型、BP神经网络模型和最小二乘支持向量机模型（least squares support vector machine,LSSVM）。结果显示，使用经过SNV预处理后的光谱进行CARS筛选得到的359个波长建立的LSSVM模型预测效果最佳，校正集的相关系数和均方根误差分别为0.937 5和0.008 6，验证集相关系数和均方根误差分别为0.831 6和0.012 0，表明可见/近红外光谱技术用于温州蜜柑的含水率检测是可行的。
Content of citrus is one of the important factors affecting storage and processing of citrus. Realtime detection of water content of citrus can guarantee the quality of citrus. The visible/near infrared spectroscopy technology,as an increasingly mature nondestructive testing method,can effectively detect the water content of citrus. The Satsuma mandarin picked from the Institute of Forest and Fruit at Wuhan Academy of Agricultural Sciences were placed naturally at room temperature for reducing the water content of the Satsuma mandarin and increasing the water content gradient of the citrus. A spectrum collection device was built with Maya2000pro as a carrier. The score is obtained by dewatering in a super electric heating constant temperature blast drying oven and calculated according to GB 5009.3-2016. The fitting accuracy of partial least squares regression (PLS) model established was compared through the use of differential processing,multivariate scattering correction,standard normal variate,SG convolution smoothing,standardization and other pretreatment methods. Results showed that SNV was the most effective preprocessing method. At the same time,a competitive adaptive reweighted sampling algorithm is used to extract characteristic wavelengths to establish a partial least square regression model,a BP neural network model and a least square support vector machine model based on water content of citrus. The results showed that the LSSVM model established with 359 wavelengths obtained by CARS screening using the spectrum after SNV preprocessing had the best prediction effect. The correlation coefficient and root mean square error of the calibration set are 0.937 5 and 0.008 6,respectively. The correlation coefficient of the verification set and the root mean square error are 0.831 6 and 0.012 0. The natural placement method at room temperature improves the water content gradient of Satsuma mandarin,thereby ensuring the adaptability of the model. It is indicated that the visible/near infrared spectroscopy technology is feasible for detecting the water content of Satsuma mandarin.