Research On Classification Of Coniferous Tree Specie Of Airborne Hyperspectral Images Based On Convolutional Neural Network

Forest is an important land resource on the earth.As the region with the richest biodiversity resources,it plays a key role in maintaining the global ecological balance and protecting species diversity.Coniferous tree species play a positive role in protecting biodiversity of forest ecosystem and promoting sustainable forest management.Accurate identification of tree species is the basis of scientific management and management of forestry resources.Traditional tree species identification methods usually use manual field investigation,which requires high experience of employees,and is time-consuming,laborious and inefficient.Taking hyperspectral images of conifers in teapot experimental forest as the object,this paper studies the classification method of conifers in Airborne Hyperspectral Images Based on convolution neural network.The research contents of this paper are as follows:(1)1-Der,SNV correction and S-G smoothing techniques are used to preprocess spectral data to eliminate the effects of scattered light,baseline drift and noise in Airborne Hyperspectral remote sensing images.Principal component analysis,normalized vegetation index and feature band fusion algorithm are used to extract features from hyperspectral data,so as to overcome the "Hughes phenomenon" of classifier performance degradation caused by the increase of hyperspectral image feature dimension.Finally,the feature extraction effect is evaluated by the feature degree of gray level co-occurrence matrix.(2)Three classical convolutional neural networks(VGG-16,Inception-v2 and Res Net50)were selected for the comparative experiment of Airborne Hyperspectral Remote Sensing of coniferous tree species.Considering the stability,training efficiency and classification accuracy of the algorithm,feature graph visualization is added to help analyze the model.The VGG-16 network is optimized by simplifying the network layer structure,organizing the convolution kernel arrangement,using the Mish activation function,using the Log Softmax classification function,introducing the global pooling layer and the Earlystooping optimizer,and adjusting the Dropout layer parameters.Sample data enhancement is adopted to solve the problem of unbalanced data of various categories.(3)The feature vector dimension after Hyperspectral Feature Extraction and different feature extraction methods are compared.According to the classification results,the principal component analysis is determined as the feature extraction method,and the principal component score is 128.The optimized VGG-16 network and the original VGG-16,Inception-v2 and Res Net50 were used to classify coniferous trees in hyperspectral images.The experimental results show that compared with the VGG-16 network before optimization,the optimized network achieves 95.34% classification accuracy,improves the accuracy by 8.47%,and shortens the training time by 84.8% in the case of small inter class gap and large intra class gap of classification samples.The optimized network model can realize hyperspectral remote sensing coniferous tree classification more accurately,improve the over fitting phenomenon,accelerate the training speed,and enhance the generalization ability.This study is of great significance to the monitoring of species diversity and the development of forestry application.