CN-120014446-B - Remote sensing image anti-cloud-fog interference change detection method based on contrast learning

CN120014446BCN 120014446 BCN120014446 BCN 120014446BCN-120014446-B

Abstract

The invention discloses a remote sensing image anti-cloud interference change detection method based on contrast learning, which comprises the steps of constructing a sample set, adopting a sample set training feature extractor to obtain a trained feature extractor, adopting the trained feature extractor, a feature processing module, a spatial attention module, a converter module, a channel attention module, a connection module and a classifier to construct a change detection network, inputting two detection images into the change detection network, wherein the two detection images are different in shooting time and same in shooting area, and outputting a prediction change graph by the change detection network. The global features and the local features of the network fusion image are detected through the change constructed by the trained feature extractor, the efficiency and the accuracy are improved by adopting the spatial attention module and the channel attention module, the rapid and accurate detection can be realized even in a remote sensing image change detection task with cloud layer interference, and the robustness of the remote sensing image change detection is improved.

Inventors

GAO KUN
YANG JIYUAN
ZHANG ZEFENG
HU BAIYANG
Hu zibo
HE YUQING
CHENG HAOBO
FENG YUNPENG

Assignees

北京理工大学

Dates

Publication Date: 20260508
Application Date: 20250117

Claims (8)

1. The method for detecting the cloud and fog interference resistance change of the remote sensing image based on contrast learning is characterized by comprising the following steps of: constructing a sample set, wherein the sample set comprises at least two remote sensing images and foggy images, the number of the remote sensing images is at least two, any two remote sensing images are different, the remote sensing images correspond to the foggy images one by one, and the foggy images are obtained by the corresponding remote sensing images; training the feature extractor by adopting the sample set to obtain a trained feature extractor; Constructing a change detection network, comprising: Constructing a sub-network, comprising: providing the trained feature extractor for receiving an input image, extracting a 32 x 32 feature map, a 32 x 64 feature map and a 32 x 128 feature map from the input image, performing up-sampling processing on the 32×32×32 feature map to obtain up-sampling features, and performing down-sampling processing on the 32×128×128 feature map to obtain down-sampling features; the input end of the spatial attention module is connected with the output end of the feature processing module, the input end of the converter module is connected with the output end of the spatial attention module, the input end of the channel attention module is connected with the output end of the converter module, and the method is used for processing the fusion feature map sequentially through the spatial attention module, the converter module and the channel attention module to obtain a final feature map; Providing two said subnetworks; the input ends of the connecting modules are respectively connected with the output ends of the channel attention modules of the two sub-networks, and are used for acquiring two final feature images, and the two final feature images are connected in pairs along the channel dimension to obtain a total image; Connecting the input end of the classifier with the output end of the connecting module, and outputting a prediction change graph according to the total graph; Providing two detection images, wherein the two detection images have different shooting times and the same shooting area, inputting the two detection images into the change detection network as two input images, wherein the input ends of the feature extractor trained in the change detection network are in one-to-one correspondence with the input images, and the change detection network outputs the corresponding prediction change map.
2. The contrast learning-based remote sensing image anti-cloud interference change detection method according to claim 1, wherein constructing the sample set comprises: the cloud layer simulation image is obtained by combining Berlin noise with the split Brownian motion; Providing at least two remote sensing images; each remote sensing image is fused with the cloud layer simulation image respectively to obtain the foggy image corresponding to the remote sensing image; all the remote sensing images and all the foggy images form the sample set.
3. The method for detecting the anti-cloud-interference change of the remote sensing image based on contrast learning according to claim 2, wherein the remote sensing image is fused with the cloud layer simulation image to obtain the foggy image corresponding to the remote sensing image, and the method is calculated according to the following mode: I(x)=J(x)t(x)+A(1-t(x)) Wherein J (x) is the remote sensing image, t (x) is a parameter representing light which is not scattered and reaches a camera, a is the cloud layer simulation image, and I (x) is the foggy image corresponding to the remote sensing image.
4. The method for detecting the change of the cloud and fog resistance of the remote sensing image based on the contrast learning according to claim 1, wherein training the feature extractor by using the sample set to obtain the trained feature extractor comprises the following steps: Providing two of said feature extractors; two images are selected randomly from the sample set, wherein the two selected images are any two remote sensing images, or the two selected images are any two foggy images, or one image is any one remote sensing image and the other image is any one foggy image in the two selected images; Inputting the two selected images into two feature extractors, wherein the feature extractors are in one-to-one correspondence with the images, and outputting feature vectors by the feature extractors according to the input images to obtain two feature vectors; Judging whether the selected two images are in a corresponding relation or not, if the two images are in the corresponding relation, setting a label value to be 0, and if the two images are not in the corresponding relation, setting the label value to be 1; Calculating a loss function according to the two feature vectors and the tag value; training the feature extractor according to the loss function.
5. The method for detecting the change of the cloud and fog resistance of the remote sensing image based on the contrast learning according to claim 4, wherein the loss function is calculated according to two eigenvectors and the tag value, and is calculated according to the following mode: loss＝(1-label)euclidean(f 1 ,f 2 ) 2 +label(1-euclidean(f 1 ,f 2 )) 2 Wherein loss is a loss value, label is the label value, euclidean (f 1 ,f 2 ) is the normalized euclidean distance between the two feature vectors, f 1 is the feature vector corresponding to one selected image, and f 2 is the feature vector corresponding to the other selected image.
6. The method for detecting the anti-cloud-interference change of the remote sensing image based on contrast learning according to claim 1, wherein the processing the fused feature map sequentially through the spatial attention module, the converter module and the channel attention module to obtain the final feature map comprises: the spatial attention module processes the fusion feature map to obtain a variable feature map; and the converter module and the channel attention module process the variable characteristic diagram to obtain the final characteristic diagram.
7. The method for detecting the anti-cloud-interference change of the remote sensing image based on contrast learning according to claim 6, wherein the spatial attention module processes the fused feature map to obtain the variable feature map, and the variable feature map is calculated according to the following mode: Wherein, the For the fused feature map, σ is the activation function, f is the convolution kernel operation, maxpool is the max pooling operation, avgpool is the average pooling operation, In order to be a two-dimensional spatial attention, For the element-by-element multiplication, And (5) the variable characteristic diagram.
8. The method for detecting the anti-cloud-interference change of the remote sensing image based on contrast learning according to claim 6, wherein the converter module and the channel attention module process the variable feature map to obtain the final feature map, and the final feature map is calculated according to the following manner: M c ＝MLP(M i ) wherein Maxpool is the maximum pooling operation, For the variable feature map, T () is the converter module processing, step-by-step addition, MLP is the multi-layer perceptual processing, For the element-by-element multiplication, For the final feature map to be described, For providing the variable feature map of one of the detected images, And providing the variable characteristic diagram of the other detected image.

Description

Remote sensing image anti-cloud-fog interference change detection method based on contrast learning Technical Field The invention relates to the field of remote sensing image change detection, in particular to a remote sensing image anti-cloud and fog interference change detection method based on contrast learning. Background The double-time-phase change detection is a key task in remote sensing, relates to comparing and identifying the change of registered remote sensing images of the same area in different time phases, and has wide application in the fields of disaster assessment, urban planning, agricultural investigation, resource management, environment monitoring and the like. Pseudo-variations are a long-standing significant problem in the field of change detection because of the presence of complex textures, seasonal variations, climate variations, and changing demands. In the remote sensing image of different phases, a phenomenon in which an unchanged area is erroneously recognized as a changed area because of other differences is called a pseudo-change. Typical pseudo-variation causes can be categorized into color variations, temporary objects, shadow shapes, and cloud interference. These spurious variation problems result in shadows and projection differences of unchanged areas being falsely detected as changed areas. Wherein spurious variations caused by thin cloud interference are a significant concern. The existing remote sensing image change detection data sets generally do not introduce the influence of thin cloud layer interference, so that the change detection methods created based on the data sets do not consider the influence of thin cloud layer interference factors. The traditional manual feature-based changing method can achieve good effects in simple scenes, but is generally poor in performance in complex scenes. Algorithms based on deep convolutional neural networks (Convolutional Neural Networks, CNN) perform better. Deep convolutional neural networks are widely used in change detection to extract discriminative local features, including classical convolutional neural networks and their extended architectures such as ResNet (residual network), UNet (image segmentation network). An algorithm based on a transducer structure effectively models global context information through the encoder-decoder architecture as compared to a pure convolutional neural network, achieves impressive results in the change detection task. But it is limited by the limitations of the transducer itself and the use of local features is limited. Algorithms have been developed that fuse CNN and transducer, such as BIT-CD (change detection model), to achieve better performance in the change detection task. However, detecting network performance based on such changes in thinking still leaves much room for improvement, especially over the problem of spurious changes due to cloud interference. Chinese patent application documents (application number: 201410441207.2, application date: 2014.09.01) disclose a remote sensing image change detection method, in which after remote sensing images before and after change are input, whether the remote sensing images have fog needs to be judged, and the fog remote sensing images are identified after defogging. The method needs a large amount of processing calculation in the process of image judgment and defogging operation, and has low processing efficiency. Based on this, it is desirable to provide a method for accurately and rapidly implementing change detection that is resistant to cloud interference. Disclosure of Invention In view of the above, the invention provides a remote sensing image anti-cloud-interference change detection method based on contrast learning. The invention provides a remote sensing image anti-cloud-interference change detection method based on contrast learning, which comprises the following steps: constructing a sample set, wherein the sample set comprises at least two remote sensing images and foggy images, the number of the remote sensing images is at least two, any two remote sensing images are different, the remote sensing images correspond to the foggy images one by one, and the foggy images are obtained by the corresponding remote sensing images; training the feature extractor by adopting the sample set to obtain a trained feature extractor; Constructing a change detection network, comprising: Constructing a sub-network, comprising: providing the trained feature extractor for receiving an input image, extracting a 32 x 32 feature map, a 32 x 64 feature map and a 32 x 128 feature map from the input image, performing up-sampling processing on the 32×32×32 feature map to obtain up-sampling features, and performing down-sampling processing on the 32×128×128 feature map to obtain down-sampling features; the input end of the spatial attention module is connected with the output end of the feature processing module, the input end of the converter modu