Research On The Improvement Of Spatial-Spectral Combination Method And Small Sample Classification For Remote Sensing Images Based Deep Learning

As one of the main research directions in the field of remote sensing,remote sensing image surface coverage classification provides important foundation support for high-level research applications such as environment,ecology and economy.Deep Learning,which has been rapidly developed in recent years can overcome the limitations of traditional remote sensing classification methods and mine deep information of remote sensing image data,which has been widely used.It has certain research value to improve the existing high-resolution remote sensing image space-spectrum combination methodby combining the richness of spectral information from the field of natural imagery based primarily on spatial information classification and remote sensing data.This paper focuses on the remote sensing spatial-spectral combined classification method based on deep learning,focusing on the following contents:(1)A solution to the problem that remote sensing images are prone to slow network convergence,training over-fitting,and low classification accuracy because of the high dimensional,small sample and high resolution of the data is proposed.Gradient Domain Guided Image Filtering is used to preserve and denoise images,reduce spectral noise and ensure spatial information;GLDM processing is used to instead of taking pixel neighborhood points as spatial information greatly eliminates redundant information and reduces data pressure on training networks;the minimum noise transform is used instead of PC A to preserve the effective spectral information while reducing the noise reduction of the image.Finally,experiments on WorldView-3,Pavia University and Indian Pines images show that the proposed method can process and extract the original remote sensing image,which can effectively avoid network redundancy information and speed up network convergence.When the number of spectral dimensions of the hyperspectral image is reduced to 0.02 of the original image,the classification accuracy is basically unchanged.(2)The Block-Pixel-DCNN model is designed to overcome the advantages and disadvantages of the previous optical spectrum combined with the depth classification model.Block-Pixel-DCNN is a two-channel deep neural network,which improves the combination of remote sensing spectral information and spatial information.It uses MLP to extract high-dimensional spectral features,DCNN to extract high-dimensional spatial information,and combines two types of high-dimensional After the feature information,continue to reconstruct the data through the network.This method not only avoids the problem of error correction between models caused by the optical spectrum decision combined with the classification strategy,but also solves the data incompatibility phenomenon easily generated by the open spectrum synchronization combined strategy.Experiments with World Spectrum-3 images using space-spectral synchronization-DCNN,space-spectrum decision-binding-DCNN,Pixel-MLP and Block-DCNN show that Block-Pixel-DCNN has the highest classification accuracy.(3)Fine-tuning training model is designed In order to solve the situation that the remote sensing training samples are too few at present.Fine-tuning can construct new model weights based on other remote sensing image depth learning classification experiences.The overall classification accuracy is 0.73 in the small sample remote sensing datasets Pavia University and Indian Pines.A remote sensing image continuous training system is constructed according to the principle of transfer learning.it can improve work efficiency and greatly save computing resources through receive any data from a fixed sensor and use it as training data to quickly update model parameters.