Research On High-resolution Remote Sensing Image Classification Based On Convolutional Neural Network

With the improvement of space and spectroscopy technology,remote sensing data exhibits clear ground object details and complex spatial structure,which can provide more refined data for remote sensing image analysis and processing.Remote sensing image classification is an important branch of remote sensing applications,and its significance is mainly manifested in the aspects of land management and planning,environmental quality monitoring and evaluation,etc.As the core of artificial intelligence,deep learning has achieved leapfrog development in recent years and has solved a series of problems in many fields.In the field of remote sensing,it has become a trend to use deep learning to learn the discriminative feature representations of remote sensing images,so as to realize the division and even fine classification of ground objects,but it also faces severe challenges,such as the multi-feature fusion effect,intra-class difference,and inter-class similarity of high spatial resolution remote sensing images,the ineffective use of the spatial and spectral characteristics,and band redundancy of hyperspectral remote sensing images,etc.As a popular model for image recognition and classification,convolutional neural network(CNN)in deep learning is capable of learning features actively,and is suitable for processing complex remote sensing data.This paper adopts CNN as the main research method and considers the characteristics of different remote sensing data and different classification problems,reasonable and effective models are established to realize the end-to-end intelligent processing of remote sensing images and improve the classification accuracy,the main research contents are as follows:(1)In order to explore the multi-level feature fusion effect of high spatial resolution remote sensing images,a two-level integrated CNN structure is established to complete the scene classification task.The designed model includes the extraction of deep-level features from remote sensing images by the transferred CNN and the extraction of middle-level features by the encoder network,so as to realize the description and extraction of multi-level features of high spatial resolution remote sensing images.Then the features of different levels are sent to the designed decoder network to complete the extraction and classification of fusion features.In addition,on the high spatial resolution remote sensing datasets——AID and RSI-CB with uneven sample distribution,adopting the optimization methods of dropout and early stop has achieved better performance than other classic CNNs.Moreover,an overall accuracy of 99.65% was achieved on the generalization experiment of hyperspectral remote sensing dataset——IP,which shows that the middle-and deep-level feature fusion classification model constructed in this paper is effective,and provides ideas for the selection of different levels feature fusion.(2)In view of the different characteristics of spectral dimensions and spatially adjacent pixels in hyperspectral remote sensing data,it leads to the problem that the spatial and spectral characteristics are used ineffectively.This paper constructs two stacked CNN models based on the spectral-spatial joint attention mechanism,S-SSA and P-SSA,to improve the classification performance of hyperspectral remote sensing image by the utilization of valid data.The proposed spectral-spatial joint attention mechanism automatically obtains the importance of spectral dimension characteristics and spatial pixel by learning,and assigns different weight coefficients to them,thereby strengthening important features and suppressing unimportant features.In addition,combined with early stop strategy and parameter verifications,the models have capability to extract more strong and discriminative features.Verification experiments were carried out on three publicly available hyperspectral remote sensing datasets,the classification performance with limited training samples surpasses the classic comparison algorithms.Among them,the overall accuracy,average accuracy and Kappa coefficient on the SA dataset exceed 99%,which proves the effectiveness of this method.(3)Aiming at the problems of spectral band redundancy and spatial spectral feature variability of hyperspectral remote sensing data,this paper constructs a 3D ImInception model based on adaptive band selection to classify hyperspectral remote sensing images on the basis of the 2D Inception structure.The dimensionality reduction based on interactive information entropy is used in the band selection of hyperspectral remote sensing data,then the low-redundant band with more abundant and discriminative information can be selected.The designed network adopts the way of nesting the network to use different kernel sizes for parallel computation.For different parallel branches,different sizes and layers of convolutions are adopted.This diversified feature extraction method can generate more flexible feature maps for different datasets.This method is applied to three commonly used hyperspectral remote sensing datasets,under the condition of limited training samples,its overall accuracy,Kappa coefficient,and average accuracy have been greatly improved,among which the overall accuracies on the PC and Botswana datasets both exceed 99%,achieving a further improvement in the classification accuracy of hyperspectral images.(4)To further verify the stability of the three types of models constructed in this paper to deal with complex data,and compare their classification performance under the same conditions,the above three types of models are applied to two airborne hyperspectral remote sensing datasets for comprehensive testing.The datasets have been publicly available in the past two years and have relatively more complex scenes.The analysis from the aspects of convergence process,parameter amount,classification accuracy,etc.shows that the three types of methods constructed in this paper all have the capability to deal with complex scenes.Among them,the P-SSA model can ensure high classification accuracy in complex scenarios,and the complexity of the model is low.In summary,this paper establishes a series of CNN classification models aimed at high-resolution remote sensing data,which have good classification performance under the condition of uneven sample distribution or limited training samples,and realize the intelligent classification of ground objects,and significantly improve the classification accuracy of high-resolution remote sensing images.In the follow-up research,according to the characteristics of different algorithms,the classification goals and focus,the appropriate classification method is selected to achieve the better remote sensing image classification.