Research On Remote Sensing Image Scene Classification Based On Deep Convolutional Neural Network

Remote sensing satellite imagery can provide high-precision,large-scale,and periodic observation data on the surface of the earth,and has been widely used in the fields of global climate change research,environmental monitoring,land and resources survey,and urban planning.How to quickly,efficiently and automatically analyze and interpret remote sensing data has become an urgent problem to be solved.Remote sensing image classification is an important step in remote sensing data processing.It is a process of judging and identifying information such as category attributes and spatial distribution characteristics（such as spatial location and area size）based on the differences in the characteristics of interested objects on remote sensing images.The traditional manual feature-based classification method is difficult to capture the rich semantic information contained in remote sensing images,which is greatly restricted in practical applications.The types of scenes that can be distinguished are relatively limited,and the generalization ability is poor.It is difficult to use Process images outside the training set.Although the method of modeling semantic description of scenes based on middle-level features improves the classification accuracy to a certain extent,the efficiency is low,and still requires a priori knowledge of manual feature extraction,lacking the flexibility to find high-level complex structural features.In recent years,deep learning has achieved remarkable results in the field of computer vision,such as image classification,face recognition,and image retrieval.This paper draws on the relevant theory and practical experience of deep convolutional neural networks in the field of computer vision images,applies them to the classification of remote sensing image scenes,and uses the ability of deep convolutional neural network models to learn the essential characteristics of data to reduce the classification of remote sensing images The work relies on manpower to achieve an end-to-end,intelligent,and automated processing effect that meets the requirements for remote sensing image classification accuracy and processing speed.The main innovative work and conclusions of this article are as follows:（1）We combine data augmentation and transfer learning strategies to classify scenes for high-resolution remote sensing images,and improve on existing shortcomings such as artificially extracted features,poor robustness,and large amount of calculation.In order to alleviate the limited availability of labeled data in the remote sensing field,we extend the remote sensing image dataset through a data augmentation strategy.In view of the high rotation variability and scale diversity of remote sensing images,the remote sensing image data set is augmented by operations such as mirroring and rotation.In addition,we use transfer learning strategies to reduce the parameters that need to be trained,to some extent,reduce the problem of overfitting,and also greatly reduce the model training time,and get better remote sensing image scene recognition results.On the UC Merced Land Use and NWPU-RESISC45₄500 datasets,the performance of 5 networks including VGG16,VGG19,ResNet50,InceptionV3,and DenseNet121 is verified,which is superior to traditional methods.（2）We combine feature fusion and multi-kernel learning support vector machines to classify scenes for high-resolution remote sensing images.Due to the unique features of high rotation variability and scale diversity of remote sensing images,we use the hand-designed Dense-SIFT feature extraction operator to perform dense rotation scale invariant feature transformation feature extraction,and PCA dimensionality reduction and Fisher coding operations are used to obtain middle-level semantic features.In view of the high expression of depth features for remote sensing images,deep convolutional neural networks are used to extract high-level semantic features of the images as global feature descriptors.Then we integrate the middle and high-level features and use multikernel learning support vector machines instead of traditional single-kernel support vector machines for scene classification.Using a multi-core learning-support vector machine can effectively avoid the need for a single-core support vector machine to choose different kernel functions,specify different parameters,and when the characteristics of the data set are heterogeneous,the effect is not ideal.The performance of the proposed algorithm is verified on the NWPU-RESISC45₄500 data set.The experimental results show that whether it is a fusion of Fisher kernel encoding middlelevel features and deep learning high-level semantic feature classification results,or a fusion of different deep learning features classification results,are better than Use single network classification results.The best classification accuracy of each group of experiments reached 91.33%,which was 1.11% higher than the best classification accuracy using fine-tuned DCNN.At the same time,the classification results of multicore learning support vector machine are also better than the classification results of traditional single-core support vector machine.