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

Remote sensing images contain information such as surface vegetation coverage,surface temperature,and urban-rural distribution information,which are of great significance for guiding macro-surface research.The remote sensing image semantic segmentation technology is to classify each pixel of the remote sensing image.It is an important and basic part in the understanding and analysis of remote sensing images,and it has extremely high research value.The latest development of convolutional neural networks has achieved solid progress in semantic segmentation of remote sensing images,but the structural design of existing convolutional neural networks does not fully take into account the data characteristics of remote sensing images,and the acquisition of network structures requires researchers to pass With continuous experimentation,the segmentation accuracy and efficiency need to be further improved.Therefore,this paper further studies the task of remote sensing image segmentation based on convolutional neural networks.The main research contents are as follows:(1)In order to solve the existing remote sensing image segmentation method based on convolutional neural network,the characteristics of the remote sensing image data are not fully considered,resulting in poor classification of small objects and unclear boundaries.This paper proposes an improved semantic segmentation neural network.This network uses hollow convolution,fully connected fusion path and pre-trained encoder to perform the semantic segmentation task of remote sensing images.The use of hollow convolution improves the segmentation of small objects such as road water,and fully connected fusion paths provide more underlying information flow through jump connections.In the training phase,the cross-card synchronization of the BN layer was realized based on the deep learning framework Tensorflow.The experimental results show that the improved neural network architecture enhances the remote sensing image segmentation results,achieves 91%classification accuracy,and improves the recognition accuracy of small objects.(2)Although the above method further improves the accuracy of remote sensing image segmentation,it requires a lot of experiments to continuously verify and the efficiency is low.Combining the importance of multi-scale representation of image information in high-resolution remote sensing image segmentation tasks,this paper further proposes a remote sensing image segmentation method based on differential multi-scale architecture search.This method expands the automatic search of network architecture to the field of remote sensing image segmentation for the first time,realizes the automatic design of remote sensing image segmentation network,liberates manpower and computing resources,and greatly improves the efficiency of remote sensing image segmentation.This method defines a feasible search space based on a priori knowledge,and continually searches the search space.Based on the gradient descent algorithm,the network architecture suitable for the remote sensing image dataset is trained and searched,and then the model is trained in the new dataset to obtain the network weights.In the experimental design part,the time spent on network automatic search and the performance of the model are evaluated in two parts.The method proposed in this paper shows that the automatic search of the architecture is three orders of magnitude lower in time and computing resources than the existing search methods.The performance of the model is better than the multi-scale architecture designed by humans,and it shows the best performance on the data set.The mloU coefficient reaches 78%.