Research On Remote Sensing Image Scene Classification Algorithm Based On Few-shot Learning

With the development of remote sensing technology remote sensing images are more and more commonly used in various fields.Previously,many studies have been conducted to solve remote sensing image scene classification tasks based on deep learning,but most of them face the problem of insufficient dataset size.To address this challenge,this paper aims to use few-shot learning methods to solve remote sensing image scene classification tasks and improve the performance by improving the models.This paper investigates the remote sensing image scene classification task based on the metric learning approach in few-shot learning,and the main contents are as follows.Firstly,the research background and significance of this topic are introduced.The application value of remote sensing images in various fields and the difficulty of remote sensing image scene classification are analyzed,and the advantages and necessity of few-shot learning method in remote sensing image scene classification are explained.At the same time,the progress and shortcomings of domestic and foreign related researches in few-shot learning are summarized and analyzed.Secondly,to address the problems of prototype bias and unreasonable weight distribution of feature dimensions in the prototype network,two improvement strategies are proposed: a sample-related attention prototype generation network is proposed to use the attention mechanism to weight the samples to generate more effective prototypes;for the analysis of similarity and difference between different categories depending on specific feature dimensions,the feature-related attention mechanism is used to give more discriminative feature dimensions to improve the classification performance.The effectiveness of the improved model is verified by experiments.Thirdly,considering that remote sensing images contain many highly complex scene things,it is necessary to use both global and local information to make classification judgments.In this paper,multi-scale feature extraction network is proposed to obtain more adequate image information and thus obtain better classification results.The effectiveness of the method is demonstrated by conducting experimental analysis in the dataset.Finally,an improved optimization method is proposed for the training process of the prototype network.For the problem of prototype validity,a prototype self-calibration strategy is proposed to improve the validity of the prototype by improving the training strategy;for the deficiency of the cross-entropy loss function,a hybrid function with fused contrast loss is proposed to solve the problems of insufficient inter-class separability and weak intra-class aggregation.The effectiveness of these optimization methods in the prototype network is proved by experimental knots.