Identification of seed maize fields from hyperspectral imagery by fusion of spectral and spatial features

Grasping the planting information of seed maize is utterly important for strengthening the macro-control of the seed industry market and ensuring the safe production of conventional maize. At present, the traditional ground survey is the main method for surveying the planting distribution of seed maize, which is inefficient and expensive. And in the existing remote sensing classification methods of seed maize, most of them use texture information, and seldom use hyperspectral information. Here, we studied the “Zhuhai-1″ satellite hyperspectral data to construct the spectral and spatial features of seed maize identification, and then used the Support Vector Machine (SVM) classifier to prepare the planting pattern map of seed maize, and finally an efficient and economical identification model for seed maize based on hyperspectral satellite images was proposed. The classification uses a multi-layer mask method to classify seed maize and conventional maize in the scene of maize planting distribution. The spectral feature extraction method of the classification model compares the two methods of Class Means Matrix Clustering Feature (CMMCF) and PCA-LDA (Principal Component Analysis −Linear Discriminant Analysis), and the extraction of spatial features uses Multiscale Extended Morphological Profile method (MEMP). Then, four feature combinations of PCA-LDA, CMMCF, PCA-LDA-MEMP and CMMCF-MEMP were constructed for classification. The classification results use overall accuracy (OA), mapping accuracy (PA), and user accuracy (UA) to evaluate the model accuracy; Fisher discriminant ratio (FR) evaluates feature separability. The results of the model show that the FR value of the texture feature in the model is higher than that of the spectral feature, and the texture feature plays an important role in the extraction of seed maize. CMMCF-MEMP-SVM (CMS) model was the best, OA reached 94.10%, PA and UA extracted from seed maize were 91.28% and 92.43%, respectively. In addition, the CMMCF-MEMP-SVM model was studied on the identification effect of seed maize under the three growth stages of seedling stage, jointing stage and milk ripeness stage, and it was found that milk ripeness stage is the best growth period for seed maize identification.