SAR image terrain classification method based on domain adaption

Abstract

The invention provides a SAR image terrain classification method based on domain adaption. By use of domain adaption algorithm, by means of prior SAR image data of a source domain, a classification model is established for adaptive application in classification of another similar yet different target domain SAR image, so that the cost of obtaining a target domain SAR image label is reduced. According to the invention, the mechanism of only reserving support vectors as the training set every time when a training set is optimized and updated can substantially reduce the redundant sample quantity of the training set and can effectively alleviate the problem that an SVM classifier is restricted by computational complexity and a storage capacity; through introduction of a recovery utilization mechanism of target domain samples unable to serve as the support vectors, the target area samples can be repeatedly and efficiently utilized so that a classification face can be continuously adjusted and corrected; and at the same time, through deletion of source domain samples unable to serve as the support vectors, the quantity of source domain training samples which do not match target domain distribution can be reduced, more accurate classification results can be obtained, and the convergence is better.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method for classifying ground objects in synthetic aperture radar (Synthetic Aperture Radar, SAR) images. Background technique [0002] Synthetic aperture radar image classification is an image processing technology that uses the gray scale, texture, shape, edge and direction of different ground object samples to determine their categories in SAR images, and divides different types of ground objects into regions. Due to the particularity of SAR imaging, SAR images are very different from optical images. For example, SAR images contain more redundant information, there is speckle noise, and SAR targets are very sensitive to azimuth angles. Therefore, how to accurately classify ground objects from SAR images has always been a difficult point in SAR image interpretation. [0003] With the continuous improvement of SAR image resolution, the target information has shown an...

AI Technical Summary

Problems solved by technology

[0003] With the continuous improvement of SAR image resolution, the target information has shown an explosive growth. The traditional SAR image supervised classification problem requires the image to be classified to have the same distribution as the source image, and relies on the use of a large number of labeled observation data to establish a classification model. However, obtaining reliable labels The difficulties of the data are (1) The field survey method is limited by the spatial scope and regional accessibility, which is expensive

At the same time, the training set obtained by field investigation usually only contains a small number of representative samples of the scene, and there may be large differences in the category features of the local area of ​​the scene, which affects the reliability and accuracy of obtaining category labels.

(2) The visual interpretation method is relatively subjective, and the selected pixels often appear in spatially adjacent groups, resulting in a highly redundant training set that may contain noise data and outlier impact data, reducing the performance of the classifier. performance

(3) Every time a SAR image is classified, its labeled samples need to be collected again, which further intensifies the classification cost

However, in DA, with the addition of useful data in the target domain to the training set, the amount of calculation and storage of SVM will increase, and SVM will be affected by the process of mapping low-dimensional linear inseparable vectors to high-dimensional linearly separable vectors. The amount of computation and storage to come is limited

Images