Variety identification model for maize seeds using hyperspectral pixel-level information combined with convolutional neural network

Abstract

As an important grain crop in China, maize has many varieties and is prone to misclassification, affecting agricultural security and food production. With the development of hyperspectral imaging and deep learning technology, identifying crop varieties using a combination of both is possible. A Convolutional Neural Network (CNN), the most representative algorithm of deep learning technology to deal with the image classification task, needs a large number of training samples in the model training process. However, obtaining a large number of hyperspectral images of maize seed samples is difficult and time consuming. Aiming at the problem of the large number of modeling samples required for the traditional method based on CNN for crop identification in hyperspectral images, a variety identification model for maize seeds based on hyperspectral pixel-level spectral information and CNN is proposed. First, hyperspectral images of different varieties of maize seeds in the range of 400-1000 nm are obtained, and 203-dimensional spectral information of all the pixels of samples is extracted. Nevertheless, the enormous amount of spectral information creates the problem of dimensional disaster and greatly increases the computational cost. Second, to reduce the dimensionality of the sample spectral information, the principal component analysis algorithm is used to reduce the spectral dimension to eight dimensions, which effectively shortens the operation time. Third, the pixel-level spectral information of the sample (i.e., the spectral information of all the pixels of the sample) is applied to the Support Vector Machine (SVM) and K-Nearest Neighbor (KNN) classification models, in addition to the CNN model. The experimental results demonstrate that for the CNN, SVM, and KNN recognition algorithms, the pixel-level spectral information models show a more stable and efficient recognition effect than the seed-level one (i.e., the average of all pixel spectral information of each sample). The seed-level information model does not fully utilize the sample pixel spectrum and spatial information, which needs a large number of modeling samples. When the number of samples used to build the classification model is the same, the CNN model has a significantly better recognition effect than the SVM and KNN models. In accordance with all the pixel-level classification results, a majority voting strategy is used to identify corn seed sample variety, and the sample recognition accuracy is up to 100% (100% refers to the identification accuracy when the numbers of samples in the modeling and test sets are 0.27 and 0.32, respectively. As the number of samples in the test set increases, the identification accuracy decreases). Lastly, the t-distribution random neighbor embedding algorithm is used to realize the visualization of output eigenvalues of CNN. The features of different maize seed varieties are clearly bounded in the visualization, which adequately verifies the validity of the species recognition model based on hyperspectral pixel-level information and CNN. In the rare case of modeling seed samples, nondestructive and efficient variety identification of maize seeds is realized, which will provide a theoretical basis for precision agriculture.