Change detection method of high resolution remote sensing image based on image fusion framework

Abstract

The invention discloses a high-resolution remote sensing image change detection method based on an image fusion framework. This method extracts high-dimensional features of bitemporal images through a pre-trained deep feature extraction network, and then puts the extracted high-dimensional features of original images into a change detection network for change detection. The network framework for change detection proposed by the present invention incorporates the multi-level high-dimensional features of the original image during the reconstruction of the change result map, so that the generated change result map has high contour consistency and internal consistency. Multi-level deep supervision is introduced into the detection network to achieve higher accuracy and recall than existing methods.

Description

[0001] field of invention [0002] The invention belongs to the field of remote sensing image processing, relates to the field of computer deep learning, and in particular relates to a high-resolution remote sensing image supervision change detection method based on an image fusion framework. Specifically, it is a remote sensing image change detection method based on deep convolutional neural network. This method extracts and fuses the features of dual-temporal remote sensing images, and realizes the detection of changed areas on the image through multi-level high-dimensional feature fusion. Background technique [0003] The interaction between the evolution of the natural environment and human behavior results in constant changes on the Earth's surface. Timely discovery and periodic attention to changes in land cover are of great significance to the harmonious coexistence between man and nature. Surface change detection based on remote sensing images is an important way and ...

AI Technical Summary

Problems solved by technology

[0004] At present, the mainstream methods for supervising change detection of high-resolution remote sensing images using deep learning are divided into two categories: one is the early fusion method, which combines two remote sensing images collected in different time periods according to the number of bands. Synthesize a piece of input image data with multiple bands, and then put the multi-band data into a neural network that receives a single data input to distinguish between changed pixels and non-changed pixels. At present, the change detection method obtained by this method The result has the highest accuracy, but due to the superimposition of multiple bands of two original images in the input data, the original image information is lacking in the process of reconstructing the change map at the back end of the network, resulting in poor quality of the generated change detection result map. Contour consistency and internal consistency; the second type is the post-fusion method, which uses two images as two inputs of a deep learning network, and uses the high-dimensional feature extraction ability of the network to realize the recognition of changing pixels. The method regards the deep learning network as a black box, and implements feature extraction and change area recognition as a whole task. However, because this method mixes feature extraction and change area recognition tasks together, the learning goal of the network is relatively vague, which leads to the network have poor generalization ability

Images