High-resolution Remote Sensing Image Classification Based On Object-oriented Deep Learning

Over the past decade,with the rapid development of high-resolution Earth observation technologies,remote sensing images have entered the era of sub-meter.High spatial resolution remote sensing images have been widely used in land use and land cover,mapping,geographic conditions monitoring,smart cities,national defense security,disaster emergency response,forestry,and agriculture.However,with the increase of the spatial resolution of remote sensing images,the ground information is highly detailed,the variance within the ground objects is increased,and the variance between different objects is reduced,which makes the spectral statistics more complicated.Meanwhile,the geometry and texture features of ground objects are more detailed,these characters make the traditional pixel-oriented processing method cannot meet the demands of high spatial resolution remote sensing image intelligent interpretation.In recent years,many researchers have improved the separability of remote sensing images in feature space through a large number of feature extraction algorithms,and effectively improved the classification accuracy of high-resolution remote sensing images.However,these features are mostly low-level features and middle-level semantic features,and cannot describe high-level information.Deep learning can obtain high-level features through multi-layer abstraction,but multiple feature extraction and abstraction blur the boundary information,and the classification result cannot describe the precise boundary of the ground object.In this paper,the object-oriented image analysis method is introduced into deep learning,and a high-resolution remote sensing image classification framework for object-oriented deep learning is proposed and applied for high-resolution remote sensing image classification.The main research contents of this paper are as follows:1.Firstly,this paper analyzes the characteristics of high-resolution remote sensing imagery and elaborates the advantages and disadvantages of existing high-resolution remote sensing image classification algorithms for the problem of high-resolution remote sensing image information extraction.Then,the development and research status of high-resolution remote sensing image segmentation and classification methods are reviewed.The basic theories and methods of object-oriented image analysis methods and deep learning are systematically introduced.Finally,the quantitative evaluation index of high-resolution remote sensing image segmentation and classification is briefly introduced.2.By analyzing the existing problems of the merged criteria of high-resolution image segmentation and the characteristics of the spectral features of the features,a high-resolution remote sensing image segmentation algorithm based on the geoscience multi-spectral index is proposed.The spectral index model can enhance the specific target features and weaken the background information,which will be beneficial to remote sensing image segmentation.The algorithm introduces multiple geoscience spectral indices into multi-scale segmentation of specific feature objects.The specific process includes:(1)using the Meanshift segmentation algorithm to obtain the initial segmentation result;(2)constructing the object adjacency graph(OAG)based on the initial segmentation result,using spectral difference,shape compactness index and multiple geoscience spectrum index to establish the merging criterion,and constructs the multi-scale segmentation result according to the merging criterion;(3)the optimal segmentation metric is determined by global features of the segmentation results.Experiments are carried out with two high-resolution remote sensing image data to verify the performance of the algorithm.Compared with the typical fractal network evolution algorithm and Meanshift segmentation algorithm,the proposed algorithm obtains higher segmentation accuracy and visual effects.3.A high-resolution remote sensing image classification algorithm based on deep learning in the object-oriented paradigm is proposed.To address the limited feature recognition ability by manual feature design and blurred feature boundary problem based on deep learning.This article will combine object oriented image analysis methods with deep learning.Firstly,the multi-scale segmentation result of the image is obtained according to the high-resolution image segmentation algorithm based on the fusion multi-spectral index.The object is used as the basic processing unit to extract the low-level features such as the shape,texture,and context of the object,and the depth convolutional neural network is input to obtain the single-scale deep feature.Then,these single-scale depth features are fused to obtain multi-scale depth features.Finally,the multi-scale depth features of the objects are used to classify high-resolution remote sensing images.The proposed method is validated by three high-resolution remote sensing image experiments and compared with other multi-scale expressions and different classification algorithms.The experimental results show that the algorithm has higher classification accuracy.