Multi-scale remote sensing image fusion method based on convolution neural network

Abstract

The invention provides a multi-scale remote sensing image fusion method based on a convolution neural network. The method comprises the following steps: firstly, a multi-scale convolution neural network fusion model conforming to the fusion characteristics of remote sensing images is constructed by utilizing the properties of the convolution neural network, wherein an image to be fused is input, and a fused image is putput; Secondly, a suitable training dataset is constructed and the fusion model is successfully trained on the dataset, thirdly, the panchromatic image Pan image is converted tothe image to be fused which is needed by the model, Fourthly, the converted approximate Pan image and multi-spectral image MS are input into the trained fusion model, and the final fusion image is obtained. The method of the invention learns an adaptive multi-scale fusion function from a large amount of data, which is not designed artificially and is more reasonable by statistical learning. The experimental results show that the multi-scale fusion method based on convolution neural network can process the remote sensing images of different satellites and different bands.

Description

Technical field [0001] The present invention belongs to the field of remote sensing image fusion. Specifically, it designs a fusion method capable of fusing pan-color image Pan image and multi-spectral image MS image together so that the fused image has both hyperspectral and high spatial resolution. Background technique [0002] In recent years, remote sensing images have been widely used in various applications, such as environmental management and detection, geological hazard prevention, precision agriculture, national defense security, and so on. Due to the limitation of satellite sensors, we can only obtain high spectral resolution multispectral image MS and high spatial resolution panchromatic image Pan. But in actual applications, we need to use both hyperspectral and high spatial resolution information at the same time. High spectral resolution is used for accurate feature classification, and high spatial resolution is used for description of feature shape and texture. T...

AI Technical Summary

Problems solved by technology

[0004] IHS, PCA, and Brovey can usually preserve the spatial information of the Pan image in the fusion image well, and the implementation is simple, but they do not consider the local differences between the Pan image and the MS image, making the final fusion image have significant spectral distortion.

In multi-scale analysis, the series of image decomposition and the filters used will have a great impact on the fusion results. Typical multi-scale analysis methods such as wavelet transform have obvious spatial information distortion in the fusion results.

Although the hybrid method combines CS and MRA, the final fusion image still has different degrees of spectral distortion and spatial structure distortion, and the fusion effect is closely related to the specific fusion method selected.

[0005] The core algorithms of the above fusion methods are basically based on artificially formulated fusion rules, and the final fusion effect is also changed by the fusion rules, and these fusion rules have a strong dependence on the fusion image itself. The same method can be used on different remote sensing images. The fusion effects obtained are not the same, and the universality is weak

Images