Remote Sensing Based Extraction Of Land Cover Using Object-oriented Classification And CNN

Land resource is the cornerstone of human existence and development.Uncontrolled and unreasonable land cover methods have caused huge damage to land resources,exacerbated the deterioration of the natural environment,and affected the sustainable development of society and economy.Understanding the status of land cover is a prerequisite for protecting land resource.In recent years,the development of satellite technology has provided satellite remote sensing images with high spatial resolution and multiple characteristic information for land cover classification.Therefore,how to obtain land cover information quickly and accurately has become a hot issue in the field of remote sensing.Pixel-based feature extraction methods cannot fully utilize the feature information of high-scoring images.Object-oriented classification is based on segmentation.Object-oriented image analysis extracts feature information,which can comprehensively use the spectral,texture,and shape features of high-scoring images.Image segmentation is the first and crucial step in object-oriented classification.The quality of segmentation directly affects the accuracy of classification.High-resolution images improve the rich feature information for object-oriented classification,but it is not that the more the number of features,the better.Highly relevant redundant features will affect the accuracy and efficiency of classification.Therefore,the feature space must be optimized during classification.Therefore,the feature space must be optimized during classification.Convolutional neural network(CNN)is the most widely used network in divinity learning.Its "convolution-pooling" structure gives it a strong self-learning ability.It can automatically extract feature information from remote sensing images without the need to manually extract features.The optimization avoids human subjective errors and improves classification efficiency.Therefore,CNN is gradually applied to remote sensing image classification.This thesis uses GF-2 image as the data source,and uses object-oriented technology and CNN to extract the land cover of a certain area in chaohu city,anhui province.The main content and conclusion are as follows:(1)Object-oriented classification is done in eCognition Developer 9.0.The image segmentation algorithm selects a multi-scale segmentation algorithm.First,experimental comparisons are performed on different combinations of shape factors and compactness to find the best combination of heterogeneity factors for different features.Then use the ESP tool to determine the approximate segmentation scale of the feature,segment the image according to the result,and finally determine the optimal segmentation scale of the feature by visual interpretation.This process avoids the blindness of scale finding,reduces the scope of scale finding,and improves the efficiency of image segmentation.After the segmentation is completed,four types of features are selected: spectral features,shape features,texture features,and custom indices,and optimization tools provided by eCognition Developer 9.0 are used to complete the optimization,avoiding feature redundancy and improving the efficiency of classification.Finally,based on the experimental results of image segmentation,a four-layer segmentation classification system is finally constructed,and in accordance with the order of the segmentation scales,the nearest neighbor supervised classification algorithm is used to complete the extraction of all features.Compared with single-level classification,the accuracy of this classification method using the multi-level classification system is higher.The former has an overall accuracy of 88.32%,the Kappa coefficient is 0.8572.The latter has an overall accuracy of 94.91%,the Kappa coefficient reaches 0.9378.(2)Classification based on CNN model was completed in Spyder4.0 and Matlab2012 software.A CNN model for land cover classification was designed.The influence of the input image block size on classification accuracy and classification effect was studied.Four different image block sizes n=5,7,9,11 are used as the input of the network,and compared with the traditional machine learning method SVM.The experimental results show that the CNN(n = 9)classification has the highest overall accuracy and Kappa coefficient,both reaching 90%.From the results of inputting different image block size classifications,when the input image block is 9,the classification accuracy and effect are the best.That is,the image block size is the key to classifying land cover with CNN.Too small patches will cause the feature information extracted by CNN will be insufficient,and the "salt and pepper phenomenon" appears,but it is not that the larger the plaque is,the better the plaque is.The excessively smooth plaque leads to excessive smoothing.Some categories that are small or linear in the image are misclassified or missed,affecting classification Precision.CNN is applied to land cover classification,which can well extract the feature information of high-resolution remote sensing images,thereby improving the classification accuracy.In the classification,the appropriate patch size is selected by combining the actual situation of the data source and the ground feature.The classification accuracy and classification Improved effect plays an important role.The research results show that: in object-oriented classification,multi-level segmentation classification has better classification effect than single-level classification.CNN can quickly and efficiently classify land cover.The classification effect is best when the model input patch size is 9.