Polarized SAR image classification method based on semi-supervised depth distance metric network

Abstract

The present invention discloses a polarized SAR image classification method based on the semi-supervised depth distance metric network, and the technical problems that the traditional depth learning only considers the non-linear relationship between the sample characteristics and the classification accuracy is not high when the number of marked samples is relatively small are solved. The method comprises the following steps: inputting to-be-classified polarized SAR image data; solving a neighboring sample of the marked sample; constructing the loss function of the semi-supervised large boundary neighbor algorithm; initializing parameters of the network; pre-training the network; carrying out fine tuning on the network; carrying out classification prediction on the unmarked samples; and outputting a classification result image and classification accuracy of the to-be-classified polarized SAR image. According to the method disclosed by the present invention, by constructing a depth distance metric network, a popular learning regular term is added to the large boundary neighbor algorithm, so that problems of the influence of insufficient marked samples on the classification accuracy and the waste of information of a large number of unmarked samples are overcome; and the characteristics learned in the method of the present invention fully depicts intrinsic attributes of the samples, and the method can be applied to the earth resources survey, military systems and other technical fields.

Description

technical field [0001] The invention belongs to the technical field of image processing, in particular to a polarization SAR image classification method, in particular to a polarization SAR image classification method based on a semi-supervised depth distance measurement network. It can be used in environmental monitoring, earth resource surveying and military systems, etc. Background technique [0002] Machine learning (Machine Learning, ML), as a subfield of computer science, starts from the research of artificial intelligence, computer learning theory and pattern recognition, and constructs an algorithm that can learn knowledge from data and predict similar data. Machine learning can learn its various attributes from the original data so as to have the ability to deal with various similar problems, which is how to make the computer automatically acquire new knowledge and new capabilities in experience learning. In the field of classification and recognition of polarimetr...

AI Technical Summary

Problems solved by technology

[0005] Traditional machine learning methods need to manually extract features and the obtained shallow features cannot fully reflect the intrinsic properties of the data

However, existing deep learning methods such as autoencoders and deep belief networks use unsupervised pre-training methods without label sample guidance. The pre-training effect is not ideal, and a large number of labeled samples are still required to backpropagate network parameters. fine-tuning

For methods such as supervised convolutional neural networks, when there are few labeled samples, the network performance cannot be guaranteed, and the classification results are often not ideal.

Images