Research On Object Based Convolutional Neural Network Using Feature Combination In High Resolution Image Classification

Land cover plays an important role in urban planning,natural resource management,environmental research,biodiversity conservation and so on.Efficient and accurate extraction of land cover classification information is of great significance to promote economic development,promote the construction of ecological civilization,and realize the harmonious prosperity of man and nature.With the rapid development of remote sensing platform and sensor technology,high-resolution remote sensing image data is increasing and continuously applied to all aspects of social and economic life,such as commercial,agricultural,forestry,animal husbandry and mining,which provides an important support for the fine classification of land cover.However,the details of high-resolution remote sensing images are amplified,feature types are more complex,and classification is more difficult,which leads to the introduction of convolutional neural networks(CNN)with strong fault tolerance,feature learning and expression ability.However,because the output features of the convolutional neural network model are highly abstract and only learn the deep features from the local image blocks generated based on pixels,the boundary of the classification results is blurred and the "salt and pepper" phenomenon is serious.At the same time,the object based image analysis method can greatly reduce the "salt and pepper" phenomenon and improve the classification performance,but it is difficult to fully excavate the rich deep features in high resolution remote sensing images,which restricts the improvement of land cover classification accuracy.Based on this,this paper proposes a land cover classification method of high-resolution remote sensing image based on object convolutional neural network using feature combination(OBCNN-UFC).On the one hand,convolutional neural networks can automatically extract high-level features that are difficult to obtain by traditional object based image analysis methods;on the other hand,object based image analysis can reduce the "salt and pepper" phenomenon in the classification results of convolutional neural networks,refine the contour of ground objects and make full use of the multi-discriminative semantic features of objects.It provides a reference for the classification and mapping of land cover in high resolution remote sensing images.This paper uses GF1,WorldView2,WorldView3,ALOS,and GF2 as data sources.Based on ENVI?Pycharm?MATLAB?eCognition?WEKA?ArcGIS and other experimental platforms,extracts the depth features of high-resolution remote sensing images by using the optimal parameter combination CNN model,and combines them with the object based feature after feature selection at the optimal segmentation scale,and inputs them to the Softmax?support vector machine(SVM)and the single hidden layer artificial neural network(ANN)classifier respectively to obtain the classification results with object shape constraints.In order to verify the effectiveness of the OBCNN-UFC method proposed in this paper,this method is compared with traditional classification methods.The study reached the following conclusions:(1)Convolutional neural network model parameter selection plays an important role.Set the model input size to 4?4,6?6,...24?24,the number of convolution + pooling layers to 1-9,and the number of feature map outputs to 14,16,...32.Comparing the overall accuracy of the image classification of the five research areas by different parameter models,it is found that the best classification performance is obtained when the input size is 18?18,the number of convolution + pooling layers is 8,and the output number of feature maps is 24,that is,the excellent CNN parameter combination.By selecting the parameters of the convolutional neural network,CNN can be used to maximize the advantages of high-resolution remote sensing image classification,make full use of depth features,and improve the recognition accuracy of complex features.(2)The selection of segmentation scale based on object image analysis is of great significance to the OBCNN-UFC method.Set the segmentation scale to 10,20,30,40,and 50,respectively.Studies have shown that the OBCNN-UFC method has significant differences in the classification accuracy of high-resolution remote sensing images at different segmentation scales.The optimal classification result is obtained when the segmentation scale is 20.Over-segmentation will improve classification accuracy.Therefore,determining the optimal segmentation scale before applying the OBCNN-UFC method proposed in this paper can further improve the classification accuracy of high-resolution remote sensing images.(3)The combination of object based image analysis and convolutional neural network model is superior to high-resolution remote sensing image land cover classification mapping.Object based convolutional neural network classification method that combines depth features and object features,integrates the advantages of both,and overcomes the shortcomings of "salt and pepper" phenomenon,inaccurate borders,inability to simultaneously use deep-level features and multi-discriminate object features,etc.Effectively improve the classification performance of high-resolution remote sensing images.(4)The choice of the classifier affects the classification accuracy of the OBCNN-UFC method.During the research,Softmax,SVM and ANN are used respectively.The results show that when using the Softmax classifier,the OBCNN-UFC classification result basically overcomes the problems of "salt and pepper" and the blurred boundary of the features,but there are still a small number of objects that appear to be misclassified and missed.When using the SVM classifier,the mapping accuracy and user accuracy have been improved,and the phenomenon of misclassification and leakage has been significantly improved.Using the ANN classifier,the accuracy of the classification results obtained by the OBCNN-UFC method has been significantly improved.The phenomenon of ground mixing has been basically resolved,and the classification boundary is accurate and clear.The classification accuracy of the water body is almost 100%,and the overall accuracy of the five study areas is 0.958,0.979,0.979,0.981,0.980,which meets the output standard of land cover classification mapping.Therefore,using the OBCNN-UFC + ANN classification method can achieve the best high-resolution remote sensing image classification performance.(5)The spatial resolution and the complexity of ground features affect the classification accuracy of the OBCNN-UFC method,but the overall classification accuracy is high.This method is universal,effective,stable,and robust for high-resolution remote sensing image classification.It provides theoretical support and method reference for classified mapping of land cover.