Research On Forage Hyperspectral Image Identification And Classification Based On Improved Convolutional Neural Network

The identification and classification of grassland forage is an important step in desertification control and digital monitoring,the effective identification of forage species can provide great help to solve the desertification of grassland.Hyperspectral imaging technology is a multi-dimensional information detection technology combining spectral technology and imaging technology,it can extract image information and spectral dimension information to improve imaging accuracy and reliability.Therefore,we try to use hyperspectral imaging technology to solve the identification and classification of grassland forage.First,we use hyperspectral equipment to take field images of grassland forage and establish a database,then we extract the hyperspectral image features of forage,identify and classify them,finally we summarize the identification and classification method suitable for the hyperspectral image of grassland forage.The main research contents and conclusions are as follows:(1)Taking field hyperspectral images of forage and establish the database.We took hyperspectral images of 10 forage species from Inner Mongolia Agricultural University and Grassland Research Institute of Chinese Academy of Agricultural Sciences.Because it takes a long time to collect hyperspectral images,there will be a lot of interference factors and redundant information,so we need to preprocess the images.We first use the improved adaptive band selection method to extract band data,then we crop the image,rotate it,standardize the data,and finally establish the image database.(2)Proposing MSRA initializing convolutional neural network.Considering the robustness and applicability of convolutional neural network,we set initialization parameters to enhance network performance,improve feature extraction ability,and achieve the purpose of improving the effect of identification and classification.The experimental result show the identification and classification accuracy of forage hyperspectral images by MSRA initialization convolutional neural network was 94.50%,which proved that MSRA had good feature extraction ability and network performance.(3)Proposing Multichannel Alexnet(MAlexnet).According to the characteristics of convolutional neural network with local perception and weight sharing,MAlexnet was proposed by combining multi-channel convolutional neural network and Alexnet.MAlexnet effectively makes use of the rich information of hyperspectral images,and can extract and fuse multi-dimensional information to mine the deep information after fusion.The experimental result show the identification and classification accuracy of MAlexnet on forage hyperspectral image was 95.64%,which achieved the purpose of improving the identification and classification effect.