基于高光谱与集成学习的单粒玉米种子水分检测模型

为建立单粒玉米种子水分含量的高精度检测模型，制备了80份不同水分含量的玉米种子样本。针对玉米种胚朝上和种胚朝下分别进行高光谱反射图像采集，每份样本取样100粒，波长范围为968.05～2 575.05 nm。采用PCA快速提取单粒种子光谱，经多元散射校正预处理后，分别采用随机森林(RF)和AdaBoost算法建立单粒种子水分检测模型，并集成两种算法特征提出基于加权策略的改进RF用于单粒种子水分含量建模。利用单粒玉米种子胚朝上的光谱信息建立的改进RF模型训练集相关系数R为0.969,训练集均方根误差(RMSEC)为0.094%,测试集R为0.881,测试集均方根误差(RMSEP)为0.404%;利用单粒玉米种子胚朝下的光谱信息建立的改进RF模型训练集R为0.966,RMSEC为0.100%,测试集R为0.793,RMSEP为0.544%。实验结果表明：改进RF的泛化能力和预测精度明显优于RF和AdaBoost算法；种胚朝上的单粒玉米种子水分含量检测模型优于种胚朝下的模型。高光谱检测技术结合集成学习算法建立的玉米种子水分检测模型预测精度高，稳健性好。

Detection Model of Moisture Content of Single Maize Seed Based on Hyperspectral Image and Ensemble Learning

In order to establish a high-precision detection model of moisture content in single maize seed, totally 80 maize seed samples with different moisture content were prepared. Hyperspectral reflection image acquisition was carried out for maize embryo up and embryo down respectively. Totally 100 grains were sampled for each sample, and the wavelength range was 968.05~2575.05nm. PCA was used to quickly extract the spectrum of a single seed. After multiple scattering correction pretreatment, the random forest (RF) and AdaBoost algorithm were used to establish the moisture content detection model of a single seed, and the characteristics of the two algorithms were integrated. An improved RF based on weighting strategy was proposed to model the moisture content of a single seed. The improved RF model was established by using the upward spectral information of single maize seed embryo. The correlation coefficient R of the training set was 0.969, the root mean square error RMSEC of the training set was 0.094%, the test set R was 0.881, and the root mean square error RMSEP of the test set was 0.404%. The improved RF model was established by using the downward spectral information of single maize seed embryo. The training set R was 0.966, RMSEC was 0.100%, the test set R was 0.793 and RMSEP was 0.544%. The experimental results showed that the generalization ability and prediction accuracy of the improved RF were significantly better than that of RF and AdaBoost algorithms. The moisture content detection model of single maize seed with seed embryo upward was better than that with seed embryo downward. The maize seed moisture detection model established by hyperspectral detection technology combined with integrated learning algorithm had high prediction accuracy and good robustness.