Remote Sensing Image Scene Classification Based On Convolutional Neural Network

With the rapid development of remote sensing observation technology,the quantity and quality of remote sensing images getting higher.Remote sensing images contain abundant ground information is used for land resource planning and environmental monitoring.However,the massive information of remote sensing images increases the time of its application.Therefore,how quickly classifying the massive image data is the key to improving the utilization efficiency of remote sensing images.Traditional image classification methods mainly rely on artificial design features to extract image features,but the application of this method is limited due to its low recognition accuracy,and sensitivity to noise and other factors.Compared with traditional methods,the convolutional neural network(CNN)has made many breakthroughs in remote sensing image scene classification.In this paper,we discussed the process of scene classification of remote sensing images by the convolutional neural network model.The main research contents are as follows:(1)For the poor classification effect of a single feature scene,we propose a remote sensing image scene classification method by fusing deep and shallow features.This method combines features of different levels to complement each other and enrich the semantic information of images.Firstly,we used the scale-invariant feature transform(SIFT)algorithm to extract the shallow local features,and then used the local feature aggregation descriptor(VLAD)algorithm to encode the shallow local features to obtain the shallow global features of the image.Moreover,the pre-trained VGG-16 model was regarded as a feature extractor to extract the convolutional layer features with scene local information(deep local features)and fully connected layer features with global information(deep global features)respectively.Finally,the characteristics of depth and shallow layers are fused as input data of ELM to conduct scene classification of remote sensing images.In addition,we applied the RSCCN7 and WHU-RS19 datasets(from open sources)to test our method,and the experimental results show that our method can describe images more accurately than the single feature method,and the classification accuracy is improved by 1.66% and 4.58%,respectively.(2)For the problems of difficulty,over-fitting,and low generalization ability in training new network models from small remote sensing image datasets,we choose the transfer learning method as the basic training method to optimize and fine-tune the pre-training model.This method introduced the batch regularization(BN)and the Leaky Relu activation function to optimize the network model structure by using Alex Net and VGG-16 pre-training models and achieved the end-to-end classification pattern.According to the actual training effect of the network model on the RSSCN7 dataset,we adjusted the activation function,learning rate,and batch size to reduce the overfitting and obtain the best classification effect.We analyzed the influence of each parameter on the performance of the network model and selected the appropriate parameters for each model through the system comparative experiments.The final convolutional neural network model has significantly improved in scene classification effect of remote sensing images.The classification accuracy of the Alex Net and VGG-16 network is 9.68% and 11.42%higher than those of the original model,which further verifies that our method can improve the scene classification effect of remote sensing images.In conclusion,we adopt the convolutional neural network model to solve the scene classification problem of remote sensing images from the perspectives of the feature extractor and fine-tune the pre-training model.On the one hand,we fused multi-level extracted features to enhance the ability of feature expression.On the other hand,we optimized and fine-tuned the pre-training model to improve the over-fitting problem.Our methods have been verified on open-source data sets and obtained a good result,which can provide a reference for remote sensing image scene classification.