Density peak-based convolutional neural network method for hyperspectral crop recognition and classification

To address issues such as high band redundancy and small differences in crop spectral features for hyperspectral remote sensing images, we proposed a density peak based convolutional neural network (DP-CNN) method for crop classification research. Firstly, a local density and distance scoring mechanism is employed to obtain an optimal subset of bands with strong representativeness and low redundancy, achieving band dimensionality reduction and optimization of crop spectral features. Subsequently, the convolutional neural network (CNN) is utilized to deeply explore the coupled features of crop spectral and spatial information, completing the crop classification research. In this paper, the AVIRIS 92AV3C hyperspectral remote sensing image is selected, and comparative analyses are conducted using all bands, the support vector machine (SVM) method, and the spectral angle mapping (SAM) method. The results show that the average overall classification accuracy of the DP-CNN model is 1.41, 16.52, and 51.75 percentage points higher than that of the method using all bands, the SVM method, and the SAM method, respectively, at maximum. Meanwhile, the DP-CNN model also demonstrates significant advantages in terms of producer accuracy and user accuracy for crop classification, which verifies the effectiveness and stability of the model and provides a new approach for hyperspectral crop identification and classification.