Research On Tree Species Recognition Of Remote Sensing Image Based On Transfer Learning

As one of the most important indicators of forest resource monitoring,tree species identification has important research value in forest resource survey.The traditional tree species recognition method uses artificial feature extraction,which is time-consuming and labor-intensive,and the recognition effect is poor.Considering the characteristics of tree species images and the advantages of convolutional neural networks in image classification,the two are used in the identification and classification of tree species.This paper proposes a classification algorithm for tree species type of convolutional neural network remote sensing images based on transfer learning.This method retains the parameters of a large data set for network model training,thus effectively solving the problem of accurate identification of small sample tree species.In order to further improve the accuracy of recognition,combined with the idea of transfer learning,a complex network model is integrated and learned.The main research contents of this article are as follows:(1)Aiming at the problems of manual extraction of features in remote sensing applications,which is time-consuming and labor-intensive,the extraction accuracy is not high,and the difficulty to obtain labeled data,a classification algorithm of remote sensing image tree type based on transfer learning is proposed.First,the three models of AlexNet,Inception-v3,and VGG-16 trained on ImageNet are used as feature extractors to extract the image features of tree species,retaining most of the parameters of the model;use the AID partial data set and the self-collected experimental forest farm data set of Northeast Forestry University to train a new fully connected layer and Softmax layer to change the number of nodes in the output layer;then introduces the Dropout function to improve the overfitting phenomenon;and finally optimizes the model through back propagation.Experiments prove that the weights learned on large data sets can be used as initial weights for the classification of tree species,so that the network training can achieve high classification accuracy using only a limited number of labeled images.(2)Based on the research of remote sensing images based on convolutional neural networks,the migration of convolutional neural network models and integrated learning are combined to design a remote sensing image classification model of integrated convolutional neural networks based on migration learning.Use three pre-trained models of AlexNet,Inception-v3 and VGG-16 to perform integration of integrated convolutional neural network model for migration learning optimization,data sampling of remote sensing images through Bagging algorithm,and use the sampled data set to classify the three bases For training.Relative majority voting method is used to integrate the trained base classifier,and finally the test set is used to test the integrated tree species classification model to obtain the classification result.The experimental results show that the overall accuracy of the integrated convolutional neural network model on the AID data set of the experimental forest farm data set of Northeast Forestry University is 98.8%,and the Kappa coefficient is 0.987,thus proving that the algorithm in this paper can accurately identify tree species and has a good Generalization performance.