Research On High Resolution Remote Sensing Image Classification Method Based On Convolutional Neural Network

With the rapid development of artificial intelligence technology in recent years,deep learning has achieved many breakthroughs in many fields.Convolution neural network has achieved outstanding performance in the field of two-dimensional image pattern recognition based on its characteristics such as local feature extraction,parameter sharing,global pooling and so on.In the field of remote sensing image processing,the research of pixel oriented automatic classification based on convolutional neural network is becoming deeper,with its high precision and strong generalization ability.but there still exists some problems.In the aspect of model construction,the structure of convolutional neural network model used for classification is seldom adjusted directly.Instead,the classical convolutional neural network model is used to extract the features of classification samples,and then combined with the traditional remote sensing image classification model,the structure adjustment and transfer learning of convolutional neural network are rarely studied or applied.At the same time,due to the diversity of remote sensing images,models are usually adjusted for the characteristics of specific remote sensing images,while domestic remote sensing images are widely used,but few automatic classification models are designed for their characteristics.In the aspect of model hyper parameter adjustment.The hyper parameter adjustment has a certain influence on the classification precision.Most of the hyper parameter adjustment can be referred to the adjustment method of the ordinary image classification task,but the adjustment method of the input scale in the remote sensing image classification is not clear.Based on the above problems,the main content of this thesis is to study the construction of classification model of remote sensing image and the influence rule of input scale on classification accuracy.1.Aiming at the characteristics of remote sensing image,taking features such as spatial resolution and bands into account,combined with the dimensions of the data input CNN model,we adjust the hyper parameter of CNN to change the dimensions of the feature maps between each layer in the model to construct the high resolution remote sensing images classification model based on convolutional neural network.We propose the model in GF-2 satellite remote sensing image classification task and to classify local area of GF-2 images with transfer learning method.Classification.The classification results are compared with classic models.The conclusion are as follows:(1)After modifying the convolution neural network model based on remote sensing image features,it can achieve higher classification accuracy in remote sensing image classification tasks.In the classification of six types of ground objects,the overall classification accuracy of 92.9% is achieved,which is higher than the traditional remote sensing image classification method and the classical convolution neural network classification model applied in our research.(2)Using the method of transfer learning can achieve good effect in classifying the local area.This can effectively solve the problem of bad performance of global classification model in local area of remote sensing image.The overall accuracy is about 85%,which is much higher than the 71% overall accuracy achieved by ordinary classification method without using transfer learning strategy.2.We study on the input scale,which is a special hyper parameter for the model of remote sensing image classification with convolutional neural networks.Based on the fully adjusted model referring to the characteristic of remote sensing image,new models are established by adding and removing some convolutional layers.Using the above model,three representative remote sensing image data sets with different spatial resolution are trained and classified under each input scale to find the input scale that can obtain the highest classification accuracy.The input scale is called the "best scale".According to the "best scale" varies with the number of convolutions layers in the model and the spatial resolution of the remote sensing image data set to be classified,the empirical rule is summarized.Thus providing empirical reference for quickly adjusting input scale in order to improve classification accuracy under similar circumstances.The conclusions are as follows:(1)The input scale has an obvious influence on the classification accuracy of the model,and there exists an input scale corresponding to the highest classification accuracy.It's possible to adjust the input scale to further improve overall accuracy.(2)In the three high resolution remote sensing images,the best scale changes with the change of the number of the convolutional layer.The more the convolution layer is,the smaller the “best scale” is,and the change of spatial resolution has little influence on the “best scale”;The influence of input scale changes on the classification accuracy can be reduced by increasing the number of convolutional layer,thus improving the stability of the model.