Classification Of Optical Remote Sensing Images Based On High-level Features

Optical remote sensing image interpretation is an important branch of remote sensing feature classification and detection.From the early stages of single feature mining,feature extraction,and feature-level classification processing,researchers have migrated ordinary optical image interpretation methods to optical remote sensing image data.To a certain extent,it improves the classification accuracy and broadens the practical application field of optical remote sensing images.However,optical remote sensing images have richer representation information than ordinary optical images.A single underlying feature is difficult to meet the diversity of real scenes.Traditional feature mining classification methods have incomplete coverage of underlying features and low feature utilization,Missing high-level edge feature capture,lack of intuitive expression,etc.In view of the above problems,a high-level feature extraction and visualization algorithm is proposed.The feasibility of the algorithm is verified by comparing and analyzing with the traditional low-level feature processing algorithm and mid-level feature engineering algorithm.The specific research contents are as follows:Aiming at the traditional low-level and mid-level feature processing algorithms,this paper adopts an improved local binary pattern feature encoding method to improve the effectiveness of features,and at the same time screens the optimal feature-level classifier to meet the image classification effect;combined with popular learning algorithms to Realize the dimensionality reduction of the underlying feature space,capture the feature information space with the highest classification value,and improve the efficiency of feature utilization to a certain extent;on the basis of the feature space fused with scale-invariant features,combined with superpixel saliency features to highlight optical remote sensing images The edge feature information of a certain extent makes up for the incomplete loss of features.In order to introduce the deconvolution neural network in the deep learning algorithm to extract high-level features,an MDCNN network based on high-level features is proposed.The multi-layer deconvolution network is used to extract the high-level features of the target optical remote sensing image,and then use the high-level features as migration The input layer of the network is trained and finally classified with a classifier.Based on this research,multi-layer pyramid deconvolution structure and soft probability deconvolution are introduced,and soft probability deconvolution is embedded in the front end of the deconvolution layer.An improved MDCNN network is proposed to effectively improve the traditional pooling The information loss caused by the process,and the spatial pyramid structure is used at the front of the classifier to retain the high-level information space to the greatest extent.The comparative test in the experimental data of Satellite,NWPU and UC Merced fully verified that the algorithm has a good effect on the multi-class classification of optical remote sensing images.From the perspective of practical application of optical remote sensing,combined with the above research results,a feasible classification algorithm for optical remote sensing images based on high-level features is proposed and implemented to realize the interpretation and classification of remote sensing images.Strong,high degree of visualization,with strong research significance.