A Multispectral Remote Sensing Image Fusion Method Based on Classification Learning

Abstract

The invention belongs to the field of image processing, and proposes a remote sensing image fusion method based on classification learning, including the following steps: S101, M 1 , M 2 and L 1 Convert to surface reflectance data, and perform image registration; S102, pair t 1 High resolution image L at all times 1 All spectral bands of all spectral bands are classified and processed to obtain a high-resolution classification image K with a total category of k; S103, calculating L 1 , M 1 , M 2 The reflectance images after data regularization respectively; S104, according to the classification result of the high-resolution classification image K, obtain the reflectance "atom" matrix of all categories of the reflectance image; S105, calculate t 2 Sparse coefficient matrix and high-resolution overcomplete dictionary under each category at time; S106, calculate target time t by category 2 The high-resolution category reflectance of each category j under , and restore to the target time t 2 High resolution regularized image under . The invention effectively maintains the spectral distribution characteristics of the image, and can be widely used in the field of remote sensing image processing.

Description

Technical field [0001] The present invention belongs to the field of image processing, and more particularly to a remote sensing image fusion method based on classification learning. Background technique [0002] Remote sensing image fusion has developed from traditional full color and multi-spectral fusion to today - empty-spectral integration, its fusion quality and geographical physical significance are effectively improved, and the technical means used in integration is also more abundant, such as based on Fusion method of main component transform, wavelet transform, etc. In recent years, as super-resolution reconstruction, compression perception and sparse learning theory and technology have become increasingly proposed by machine learning methods, in particular, the fusion method based on remote sensing inversion is preliminarily presented, such as based on single data pair And the sparse learning fusion algorithm for dual data, the main idea is to perform dictionary-based ...

AI Technical Summary

Problems solved by technology

The advantage of this type of method is that its fusion accuracy is higher than other existing fusion methods for land cover areas dominated by phenological changes, but the disadvantage is that its fusion accuracy and stability are poor for land cover areas dominated by type changes, especially is for using a single data pair (known research area t 1 High and low resolution images at time instant and t 2 The low-resolution image at the moment, use the fusion method to find t 2 This problem is particularly prominent in the sparse learning method of high-resolution image data at all times

The main reason for this phenomenon is that a remote sensing image generally contains multiple object categories, and the radiation spectral attributes of each object category show differences with time and space. Image learning and training make the resulting "over-complete" sparse dictionary unrepresentative, and this problem cannot be effectively solved by adjusting training parameters and training termination conditions

Images