Research On Small Sample Remote Sensing Image Classification Based On Meta-learning

With the progress of remote sensing technology,the actual demand for accurate classification of remote sensing image data sets has increased significantly,which has become a hot research direction in the field of remote sensing technology.It is mostly used in disaster monitoring,resource planning and other fields.Traditional machine learning classification algorithms are difficult to describe the rich spatial semantic features of remote sensing images,resulting in low classification accuracy.At the same time,due to the cumbersome processing procedures of remote sensing images,it is more difficult for researchers to obtain large-scale remote sensing image data sets.The success of traditional deep learning algorithms is inseparable from data driving.The deep learning algorithm learns the mapping relationship between input and output through a large amount of data.The deep learning algorithm learns the mapping relationship between input and output through a large amount of data,while the remote sensing data set is difficult to provide a large number of samples to help the neural network complete the correction of the mapping relationship by using the directional propagation algorithm.Therefore,this paper takes the few-shot remote sensing image classification as the research object,and studies how to quickly learn and generalize the classification model from few-shot remote sensing images through meta-learning and metric learning.The main contents are as follows:(1)In order to overcome the poor performance of traditional machine learning model in few-shot remote sensing image classification and enhance the autonomous learning and rapid generalization ability of the model,a few-shot classification algorithm based on variational autoencoder and optimizer meta-learning is proposed in this paper.A meta-learning algorithm based on long and short-term memory neural network was designed to minimize the lower bound of the loss by fitting the network parameter update mode with the gating structure.It has the mechanism of automatically learning the parameter update mode of classifier.Compared with the traditional method,it can effectively expand the search space of the optimization algorithm.At the same time,considering the cross-category knowledge transfer and training time of samples,using the idea of transfer learning,the data of different categories are mapped to the same feature space,and the classifier trained by representation is meta-trained,so that the classifier can better grasp the overall characteristics of the category and accelerate the training process of meta-learning.The experimental results verify the excellent performance of the algorithm and provide a solution for the classification of few-shot remote sensing image data sets.(2)Considering the cross-category difference of few-shot remote sensing image data,this paper designed a multi-scale and multi-measure network to improve the generalization and representation ability of the model and reduce the redundancy of the model.First use of parallel multi-branch feature extraction was used to extract the multi-scale features,to grasp the image based on multi-scale feature space semantic information and scene semantic information,secondly using attention mechanism network for multi-scale feature fusion,and then the support set Mapping and cascading with the attention feature of the target set to obtain the relationship feature between the support set and the target set,and finally the relationship feature is passed into the CNN measurer to get the result.