Feature Representation Learning Of Remotesensing Image Based On Few Shot Learning

Remote sensing technology is widely used in geology,agriculture,ocean and other fields,and has become the main technology for analyzing macro geographic information.Over time,earth’s surface covering in the same area is likely to change,and these changes are often the things that we focus on.The change detection of remote sensing images refers to that,for remote sensing images and related data in the same area in different periods,image processing and mathematical model technology are used to compare,analyze and study the changes between images.This direction is the focus of remote sensing research at present.With the rise of deep learning,the change detection method based on deep learning has become a new research focus in recent years.However,most of the existing deep learning change detection methods are supervised,and there are two main problems with such supervised methods: first,deep networks cannot directly transfer and detect heterogeneous data.This means that training is needed for different data sets,and one data set in the change detection task is two images,which not only adds complexity to the change detection task,but also fails to take advantage of the feature commonality between similar images.Secondly,the network performance used for detection depends on the number of training samples.Therefore,the performance of supervised methods inevitably depends on the performance of classical methods in pre-classification.The unsupervised change detection methods does not use the pre-classification,but they have two main problems:one is that it needs more iterations to update,and many parameters need to be set manually.Second,the effect of unsupervised method is generally not as good as that of supervised method.Therefore,aiming to address above problems,this paper proposes three novel change detection methods,which greatly improve the change detection process and at the same time improve the detection accuracy by using the commonness of data characteristics.1.Aiming at characteristics of two sets of input images in the change detection task,a new SAR image change detection method based on complex value network is proposed,and experiments are carried out on three different SAR image data sets.The main idea of this method is that the original image in the conventional supervised change detection process is directly combined by channels,and the method is improved to construct complex value data,and the corresponding complex value network is designed for newly constructed complex value data to complete feature extraction and detection.The trained complex value network can accomplish final detection by following process of supervised change detection algorithm.Experimental results show that the algorithm can improve detection accuracy to some extent.2.Existing change detection algorithm can achieve good migration between different data sets,and both supervised and unsupervised method can make good balance in precision and artificial participation,so we put forward a novel based on measured yuan change detection model of learning,called CDMN,the main idea of this method are as follows:the multi-source heterogeneous image change detection as a meta learning problems.First,inspired by active learning,we designed a new method of training sample selection,k-PLR,to screen a small number of highly reliable and representative samples.Then,we build a meta-learning network composed of convolution module and graph convolution module,and use meta-learning training strategy to complete the mapping and feature measurement of samples.Furthermore,the trained models can be migrated to other change detection data sets for direct testing,and the samples of test support set required for testing can be well extracted using ELR.In addition,if the training data set is the hyperspectral image data with low noise interference,the trained model migration can even improve the accuracy of the SAR image data set with high noise detection.3.Based on the problem that test support samples need to be set when meta-learning is used on different data sets,we combine clustering algorithm with meta-learning to improve the robustness and accuracy of change detection.Meanwhile,the application scope of the method is extend to multi-class change detection.The main idea of this method is: the measurement of learning model in the test needs to set a small amount of labeled samples as test support set,but in change detection task there is no label.So we using clustering module instead of original measurement module in testing,making model using different structure in the training and testing.This method not only combines the original measurement module characteristics with clustering module characteristics to improve the accuracy,but also extends application scope of the model from two-class detection to multi-class detection.Experimental results show that our framework can not only achieve the best results in 2 classes detection problems,but also achieve comparable results in multi-class detection problems.