High-resolution Remote Sensing Image Scene Classification Based On Adaptive Deep Belief Network

With the continuous development of remote sensing technology,the spatial resolution of remote sensing images has been significantly improved.The clear and rich details of the objects in the images have brought great challenges to the scene classification of high-resolution remote sensing images.High-resolution remote sensing image scene classification is an important part of remote sensing image interpretation.The traditional classification method cannot describe high-level semantic information of complex scenes due to the diversity of features,the variability of features,and the diversity of the same object in the image scene.In recent years,the emergence of deep learning algorithms provides an effective way to extract high-level intrinsic features of images.The key to achieving high-resolution remote sensing image scene classification is image feature extraction and feature select,and image texture features are the most commonly used in remote sensing image processing.Aiming at the problem that the classification accuracy of image scene is not high due to the small number of samples,this paper proposes a high resolution remote sensing image scene classification method based on adaptive deep belief network model(ADBN),which combines the dual-tree complex wavelet transform,adaptive step size technology and deep belief network.The main research contents of this paper are as follows:(1)A new adaptive learning rate method is proposed,which not only dynamically adjusts the learning rate according to the change of reconstruction error,but also considers the influence of the variation of reconstruction error on the learning rate.Based on this,the adaptive deep belief network is constructed.The experimental analysis on the MNIST dataset shows that the model can converge faster and improve the classification accuracy of the data set.(2)The texture features of the wavelet domain of high resolution remote sensing image are extracted as feature vectors for scene classification.In this paper,the image is decomposed into multiple sub-band images based on the dual-tree complex wavelet transform method.The feature vector of the high-frequency sub-band and low-frequency sub-band of the image are constructed by the generalized Gamma density model and the local binary pattern.The experimental results show that the proposed method can achieve better results in high-resolution remote sensing image scene classification.(3)This paper also studies the classification effect of texture feature based wavelet domain and ADBN model in high resolution remote sensing image scene classification under different spatial resolutions.The optimal network layer number and hidden layer nodes are selected by comparing classification accuracy,Kappa coefficient and classification duration.Also,the experimental results show that the proposed model has similar applicability and robustness in high-resolution remote sensing image scenes with different spatial resolutions by comparison with traditional methods.