Research On Scene Classification Of Remote Sensing Images Based On Deep Learning

With the rapid development of remote sensing information technology,most spacecraft cameras have the ability to capture high-resolution remote sensing scene images,providing rich ground information.In order to efficiently obtain effective information from remote sensing scene images,image classification technology has attracted much attention.In recent years,due to the rapid development of deep learning,convolutional neural networks have developed rapidly in image classification research and gradually replaced traditional classification algorithms.However,remote sensing scene images usually come from aerial equipment,the shooting angle is large,and the dataset usually has the characteristics of complex terrain information,resulting in the inability of conventional convolutional neural networks to effectively extract advanced semantic information.Therefore,from the perspectives of deep learning and optimized support vector machine(SVM),this paper focuses on the benchmark dataset of remote sensing images in view of the two major challenges of remote sensing scene images,namely intra-class diversity,inter-class similarity,and slow optimization speed and low classification accuracy of traditional SVMs.(1)Aiming at the problems of traditional manual features relying on manual design,weak ability to express features,feature fusion increasing model complexity and computation,and poor performance of deep learning when lacking a large number of data labels,a remote sensing image scene classification method based on deep learning feature fusion(GL-LBP-CNN)is proposed.The proposed shallow joint feature effectively integrates the overall and local texture information of the image,and integrates it with the depth semantic information to obtain more discriminating features.The model does not affect the classification accuracy to a certain extent,and improves the classification efficiency.Experimental results in UC Merced and RSSCN7 datasets show that this method has better classification performance than traditional classification methods and existing deep learning methods.(2)Aiming at the problems of local optimal solution and slow optimization speed in SVM parameter optimization of remote sensing scene images and traditional optimization algorithms under small sample datasets,a method based on heterogeneous joint feature fusion and lion group optimization SVM(Lions optimization algorithm SVM,LSO-SVM)is proposed.The proposed heterogeneous feature fusion method based on transfer learning can enhance the robustness and generalization ability of features,increase the diversity of image features,and reduce the feature dimension.Furthermore,the LSO-SVM algorithm is proposed for SVM optimization,which significantly improves the classification performance compared with the traditional optimized SVM algorithm.Experimental results in the UC Merced and RSSCN7 datasets show that this method outperforms the vast majority of machine learning and deep learning-based methods.(3)Aiming at the great challenges of intra-class diversity and inter-class similarity of remote sensing images and the poor performance of attention mechanism at the end of the embedding network,a Moblie Netv2 model based on attention convolution module(ACM)is proposed.The proposed ACM module adopts a combination of spatial attention,channel attention and 1*1 convolution.The two activation functions of Hard-Sigmoid and Hard-Swish are introduced to effectively reduce the calculation amount of the model.On the RSSCN7 and RSOD datasets,better classification performance is achieved at the end of the Convolutional Neural Network(CNN)than other attention mechanisms.(4)Aiming at the challenges of bias of advanced semantic information and scene features in remote sensing scene category labels,and diversification and granularization of data sets,a multi-granularity homologous feature fusion strategy(MGCNN)based on this paper is proposed.The proposed multi-granularity residual information method can alleviate the inconsistency of scene characteristics and semantic tags while retaining the originally lost information.In the feature fusion stage,the granularity and the corresponding granular residual information are fused by bilinear pool,and the second-order semantic information after fusion is richer than the traditional connection method.And through the Sparrow optimization algorithm-SVM(SSA-SVM)to obtain higher classification performance.Experimental results in the UC Merced and AID datasets show that this method not only outperforms most deep learning methods,but also achieves a high level with a very small number of training samples.In summary,starting from the image classification technology of deep learning and optimized SVM algorithm,this paper carries out a series of studies on the challenges of remote sensing scene images,and improves its classification performance and generalization ability in different research problems.