Small sample city remote sensing image information extraction method based on domain conversion and pseudo labels

Abstract

The invention provides a small sample city remote sensing image information extraction method based on domain conversion and pseudo labels. The method comprises the following steps of: constructing a source space-target space domain conversion network, inputting labeled data of a source space into a generation network 1, and converting the labeled data into an unlabeled small sample data space; inputting the obtained small sample space image and the real small sample data into a judgment network 1; constructing a target space-source space domain restoration network, wherein a generation network 2 restores the image after domain conversion to an original image space, and a judgment network 2 realizes that a source space image generated by the judgment network 2 is the same as real source space data in distribution; constructing a semi-supervised coding network, inputting the image after domain conversion and the unlabeled image of the small sample space to obtain a corresponding implicit vector, and obtaining a pseudo label through a Gaussian process; constructing a semi-supervised decoding network, and inputting the implicit vector to obtain a corresponding prediction result; and optimizing the constructed network model, and obtaining city typical element classification in the small sample city remote sensing image based on an optimization result.

Description

technical field [0001] The invention belongs to the field of remote sensing image information extraction, and relates to a small-sample urban remote sensing image information extraction method based on domain conversion and pseudo-label generation. Background technique [0002] At present, the sustainable development of cities is facing severe challenges, and "urban diseases" such as urban waterlogging, urban heat island, and urban ecological function degradation have become global problems that need to be solved urgently. The implementation of the "High Score" project provides usable data for remote sensing to obtain global urban geographic information. However, limited by the level of intelligent interpretation of remote sensing images, it is an urgent need and difficult problem for urban information extraction to interpret remote sensing images and extract various ground features to meet the needs of urban fine management. [0003] In recent years, with the development o...

AI Technical Summary

Problems solved by technology

However, the method training based on deep learning relies on a large number of high-quality labeled samples. Limited by factors such as geographical location and weather, there are only a small number of samples available in some cities. The characteristics of the ground features are quite different, and the phenomenon of "same object with different spectrum" is more obvious

Therefore, existing deep learning-based information extraction models have low generalization ability in remote sensing data

Images