Research And Application Of Deep Convolutional Neural Networks In Remote Sensing Image Classification

Remote sensing image classification is a crucial link in remote sensing digital processing.Through remote sensing image classification,remote sensing images are converted into terrain classification information that can be understood and processed by computers,which can provide support for high-level applications.Traditional machine learning methods used in remote sensing image classification require heavy feature analysis and extraction.Not only do they have relatively stringent requirements for data representation,but each method has its own special application limitations,and its limitations are obvious.The recent development of deep learning technology can greatly overcome this limitation.Deep convolutional neural networks(DCNNs)are very well-developed deep learning models in the field of image processing.Because remote sensing is closely related to image processing technology,it draws on the successful experience of DCNNs in image processing and applies this technique to remote sensing image classification research.It is highly feasible and valuable for research.This article mainly studies the application of DCNNs in remote sensing image classification,and focuses on the following:1)Study the principle of traditional remote sensing image classification methods.The advantages and disadvantages of supervised and unsupervised algorithms were analyzed and summarized in order to compare them with DCNNs.The principle of classification accuracy evaluation and the advantages and disadvantages of several methods are studied,and theoretical guidance is provided for rational use of these methods in experiments.2)This paper studies and analyzes artificial neural networks and DCNNs from the perspective of mathematical principles and network models.Focused on the key technologies of DCNNs.Introduced models such as LeNet,AlexNet,GoogleNet,R-CNN and FCN,as well as the practice of these networks in image classification and semantic segmentation.3)This paper studies and designs a complex domain DCNNs model that extends DCNNs from the real domain to the complex domain.Based on the idea of patch,the classification network is applied to remote sensing image classification similar to image semantic segmentation.This paper uses this model to classify polarimetric SAR images and obtains ideal results.4)This paper studies and designs an end-to-end DCNN model.This model is based on the idea of full convolution,using a symmetrical structure,can integrate position information and feature information,and local features and macro features to achieve high-precision remote sensing image classification results.The model adapts to multi-spectral and different image size inputs,enabling end-to-end intelligent processing.The theoretical research and experimental results show that DCNNs can not only be applied to remote sensing image classification,but also have excellent classification effect.DCNNs technology is still in rapid development,and it continues to deeply research this technology,and applies it to the field of remote sensing data processing,which can bring about qualitative changes in the development of remote sensing data processing to the direction of intelligence.It is foreseeable that DCNNs have potential and huge application value in the field of remote sensing.