Study On Automatic Evaluation Method Of Land Cover Classification Based On Convolutional Neural Network

Land cover classification is a compound data of ground objects with different features,which are different from the purpose of land use,and the land cover classification tend to describe the natural properties of the ground objects.Land cover classification has become an important basic data in many areas such as geographical condition monitoring,land resource management and so on.With the increasing demand of the remote sensing image surface coverage classification data,the large amount of data,the fast updating cycle and the limited spectral information of the image source have limited its large-scale automation production and evaluation.Correspondingly,in the actual quality evaluation process,it still depends on manual visual interpretation.At present,the demand for automatic quality assessment of remote sensing image classification products is very urgent in national surveying and mapping projects.Driven by actual demand,domestic related projects have carried out research on the automatic quality evaluation technology of geographic information products.In the study of automatic quality evaluation of remote sensing image classification products,the ideas of existing items are generally pixel based classification.In fact,the existing classified products are generally produced by artificial visual interpretation,so there will be a low precision based inspection of high precision products in the application.The convolution neural network,as a computer vision method,can offer the classification results of multiple scales directly through the image data as a whole,which provides the basis of constructing the automatic classification quality evaluation of the land cover.In this paper,a series of research and comparative experiments are carried out based on the theories of deep learning,convolution neural network and land cover classification.The main research and work carried out include the following aspects:(1)Compiling automatic image acquisition program and constructing remote sensing image dataset suitable for training and validation of convolution neural network.On this basis,the effectiveness of convolution neural network in automatic recognition of remote sensing image land cover classification is verified.Subsequently,a comparative experiment was conducted through the truth data set,and the network models,hyper parameters and algorithm combinations were compared.(2)Based on the survey data of the geographical conditions,the structure of the convolution neural network model and the pre-trained model for the effective classification of the satellite and aerial images within the range of 0.5 to 2 meter pixel resolution are trained.(3)Based on the convolution neural network method,an automatic quantitative evaluation method for land cover classification is proposed.Taking the map as the basic unit of evaluation,the evaluation effectiveness of single class plaques and the evaluation strategy of mixed patches are studied and verified by experiments.Finally,combined with the specific remote sensing data,the method is verified by experiments and combined with the actual effect to improve the evaluation method.This paper preliminarily forms a technical method for the automatic quality evaluation of the DOM-DLG sets and the classification data of the land cover,and realizes the screening of the suspicious image spots and the quantitative evaluation of the entire image.