Research On Rapid Classification And Accuracy Evaluation Methods For Remote Sensing Images

The rapid development of Earth observation technology has provided massive remote sensing data with high precision,high timeliness,and large area coverage in various fields.How to quickly extract useful information in massive remote sensing data and ease the status quo of "big data while small knowledge" has become one of the key issues that restrict the application of remote sensing data.In recent years,the deep learning method represented by "deep convolutional neural network" has made great breakthroughs in the fields of image recognition and segmentation,and it has also been applied to the classification of land objects based on remote sensing images.However,remote sensing images have multi-spectral characteristics and spatial location characteristics.The training of deep convolutional neural networks based on remote sensing images has the problems of low computational efficiency and lack of spatial information.At the same time,the existing accuracy evaluation method of remote sensing image classification results has the problems of high information redundancy and low evaluation efficiency,due to the spatial correlation of remote sensing images.Therefore,for the purpose of rapid classification and accuracy evaluation of remote sensing images,this paper studies the fast classification method for remote sensing image based on advanced deep convolutional neural network and the method of rapid evaluation of remote sensing image classification results based on gray level co-occurrence matrix.The main research content includes:(1)An improved depth convolutional neural network model for remote sensing image classification is proposedThe depth convolutional neural network model for remote sensing image classification proposed in this paper is mainly improved in the following two aspects:Based on the spatial location characteristics of remote sensing images,the remote sensing image trained by depth convolutional neural network model is "banded" :The spatial location features of the images are converted into "longitude band" and "latitude band".And these two bands are combined with other bands for training,so as to improve the recognition of spatial location features of remote sensing image by the deep convolutional neural network model for classification.Based on the spectral features of remote sensing images,a bottleneck unit is set in the deep convolutional neural network to realize the dimensionality reduction of the input image and the feature maps;By the group convolution,reducing the calculation of convolution in the training process;and the shuffle structure is constructed to promote the feature extraction ability of the neural networks in the group convolution stage.(2)An accuracy evaluation method for remote sensing image classification based on gray level co-occurrence matrix is proposedFor the classification results of remote sensing images,this paper proposes a method based on gray level co-occurrence matrix for the accuracy evaluation of remote sensing image classification results.The method mainly solves the following two questions: How to determine the sample size and how to determine the spatial distribution of the sample unit for the accuracy evaluation on the purpose of reducing redundancy of information between samples and improve the evaluation efficiencyFor the problem of how much sample size is extracted for the accuracy evaluation of remote sensing image classification results,this paper quantifies the spatial correlation between pixels of remote sensing image based on the gray level co-occurrence matrix,and establishes the mathematical relationship between batch size and sample size of the classification results through spatial correlation.For the problem of how to arrange spatial distribution of sample units in the accuracy evaluation of remote sensing image classification results,this paper optimizes the step size design of systematic sampling by quantifying the spatial correlation of the images;this gives a rapid accuracy evaluation method for remote sensing image classification results.Taking islands classification extraction and accuracy evaluation based on remote sensing images as an example,the rapid classification and accuracy evaluation methods proposed in this paper were verified.First,based on the improved depth convolutional neural network model for remote sensing image classification,1500 islands remote sensing image classification data set were trained and built a classification model of islands remote sensing image.Compared with the conventional model,the time consumption of proposed model was reduced to 1/18.Through the accuracy evaluation method based on the gray level co-occurrence matrix,the accuracy of multiple sea island classification results was evaluated with less sample size.The evaluation results show that the classification accuracy of these islands is over 90%.The Experiment show that the proposed classification method based on improved depth convolutional neural network for remote sensing images can quickly train classification models and achieve rapid classification of remote sensing images;the proposed classification accuracy evaluation methods based on gray level co-occurrence matrix can effectively reduce the redundancy between evaluation samples and improve evaluation efficiency under the premise of ensuring classification accuracy.