CN-122023838-A - Remote sensing image unsupervised change detection method and system based on structural sensing and noise enhancement

CN122023838ACN 122023838 ACN122023838 ACN 122023838ACN-122023838-A

Abstract

The invention discloses an unsupervised change detection method and system for a remote sensing image based on structural sensing and noise enhancement. According to the method, the shared characteristic representation is extracted from the double-phase remote sensing image, the sensing capability of the structural sensing contrast learning mechanism to the real geographic structural change is enhanced, the noise disturbance consistency constraint is designed to avoid the optimization shortcut problem, and the frequency attention decoding mechanism is adopted to finely describe the boundary of the change area. The system comprises a preprocessing module, a feature encoding module, a structure sensing module, a noise disturbance module, a frequency attention decoding module and an unsupervised optimization module. The invention effectively solves the problems of optimization shortcuts, insufficient semantic representation capability, imprecise boundary depiction and the like in the existing unsupervised change detection method without manual labeling, improves the accuracy and robustness of change detection, and is suitable for the fields of urban expansion monitoring, disaster assessment, environmental change analysis and the like.

Inventors

WANG NAN
WU QINGXI
TAO RAN

Assignees

北京理工大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. An unsupervised change detection method of a remote sensing image based on structural sensing and noise enhancement is characterized by comprising the following steps: S1, inputting double-phase remote sensing image And Normalized processing is carried out, and the network is encoded through the characteristics of sharing weight Extracting multi-scale feature representations And ; S2, representing the extracted features And Performing structure perception feature construction through a structure disturbance operator Generating post-structural-disturbance features And ; S3, representing the characteristics And Applying random noise perturbation operators Generating noise disturbance features And And decodes the network through the change Calculating the change response corresponding to the original characteristic and the noise disturbance characteristic And ; S4, in the process of variable decoding, carrying out frequency decomposition on the characteristics, separating low-frequency components and high-frequency components, carrying out self-adaptive weighted fusion on different frequency components through a frequency attention mechanism, and generating a variable response diagram ; S5, constructing a structure-containing perception contrast loss Loss of noise disturbance consistency And change response canonical constraint loss Is a joint loss function of (2) Performing unsupervised joint optimization on the feature coding network, the structure sensing module and the change decoding network; S6, inputting the double-time-phase remote sensing image to be detected into the trained model, and outputting a change detection result.
2. The method for unsupervised change detection of remote sensing image according to claim 1, wherein the structural perceptual contrast loss By characteristic projection functions Mapping the features to the embedded space, so that feature pairs with consistent structures are kept close in the embedded space, and feature pairs with different structures are kept distinguished, wherein the specific expression is as follows: ; Wherein, the , , Respectively representing the first time phase characteristic, the second time phase characteristic and the projection representation of the structural perturbed characteristic in the embedding space, Representing a feature similarity measure function, Is a temperature coefficient.
3. The method for unsupervised change detection of remote sensing image according to claim 1, wherein the noise disturbance consistency is lost Defined as the response of the change corresponding to the original characteristic and the noise disturbance characteristic And Between (a) and (b) Norm distance: The method is used for restraining the model to keep stable change discrimination capability when facing feature disturbance, and avoids optimization shortcut problems.
4. The method for unsupervised change detection of remote sensing image according to claim 1, wherein the change response is a regular constraint loss By using Norms: the method is used for enabling the change detection results to be more concentrated in spatial distribution, and accords with the characteristic that a change area in a remote sensing scene is generally sparse.
5. The method for unsupervised change detection of remote sensing image according to claim 1, wherein the joint loss function Expressed as: ; Wherein the method comprises the steps of And To balance the weight coefficients of the loss terms.
6. The method for unsupervised change detection of remote sensing image according to claim 1, wherein the structure perturbation operator The method comprises the steps of performing Patch segmentation on the feature map, and performing rearrangement, shielding or neighborhood disturbance operations on different patches to simulate potential structural changes.
7. A remote sensing image unsupervised change detection system based on structural sensing and noise enhancement for implementing the remote sensing image unsupervised change detection method according to any one of claims 1 to 6, characterized in that the system comprises: The preprocessing module is used for carrying out normalization processing on the input double-phase remote sensing image; the feature coding module adopts a convolutional neural network with shared weight and is used for extracting multi-scale feature representation of the double-phase image; The structure perception module is used for carrying out structural disturbance on the characteristics and constructing structural perception contrast learning constraint; the noise disturbance module is used for applying random noise to the characteristics and establishing consistency constraint; The frequency attention decoding module is used for carrying out frequency decomposition on the characteristics, adaptively fusing different frequency components through an attention mechanism and generating a change response diagram; The unsupervised optimization module is used for constructing a multi-constraint joint loss function and performing end-to-end training on the whole network; And the change detection output module is used for inputting the double-phase remote sensing image to be detected into the model after training and outputting a change detection result.
8. The remote sensing image unsupervised change detection system according to claim 7, wherein said frequency attention decoding module comprises: a frequency decomposition unit for converting the input features into a frequency domain by fourier transform or wavelet decomposition and separating a low frequency component and a high frequency component; An attention weight calculation unit, configured to calculate corresponding attention weight matrices based on statistical properties of the low frequency component and the high frequency component, where a weight of the low frequency component reflects a global structural importance, and a weight of the high frequency component reflects an edge detail importance; The feature fusion unit is used for carrying out weighted fusion on the low-frequency component and the high-frequency component according to the calculated attention weight to generate enhanced feature representation, wherein the low-frequency component is used for describing global structure information, the high-frequency component is used for describing edge and detail changes, and the boundary features of the change area are enhanced through the weighted fusion.
9. The remote sensing image unsupervised change detection system according to claim 7, wherein the joint loss function constructed by the unsupervised optimization module comprises structural perception contrast loss, noise disturbance consistency loss and change response regular constraint loss, and the optimization of network parameters is realized by minimizing the joint loss function, so that the whole training process does not need manual labeling data.
10. The remote sensing image unsupervised change detection system according to claim 7, further comprising a change region post-processing module for performing threshold segmentation, morphological operation or connected domain analysis on the output change response map to generate a final binary change detection result.

Description

Remote sensing image unsupervised change detection method and system based on structural sensing and noise enhancement Technical Field The invention relates to the technical field of intelligent processing of remote sensing images, in particular to an unsupervised change detection method and system for remote sensing images based on structural sensing and noise enhancement. Background The change detection is a basic and key task in remote sensing image analysis, and aims to identify and position the change condition of the earth surface coverage or the ground feature state by comparing remote sensing images acquired by the same area at different moments. The technology has important application value in the fields of urban expansion monitoring, infrastructure updating, disaster assessment, environmental change analysis, agricultural monitoring, military reconnaissance and the like. Along with the development of remote sensing technology and the improvement of satellite data acquisition capability, efficient and accurate extraction of change information from massive remote sensing data has become a focus of common attention in academia and industry. The existing change detection method can be mainly divided into two major categories, namely a supervised method and an unsupervised method. The supervised variation detection method generally relies on a large number of manually labeled training samples, and learns the distinguishing characteristics between the varied and non-varied regions through training the deep neural network. The method can obtain higher precision on a specific data set, but the application effect of the method is seriously dependent on the acquisition of high-quality labeling data, the change labeling of the remote sensing image generally requires professional field knowledge, the labeling cost is high, the period is long, the generalization capability of the model on data acquired in different areas, different sensors or different seasons is limited, and the wide deployment of the model in practical application is limited. The unsupervised change detection method aims at getting rid of dependence on manual labeling, and gradually becomes a research hotspot in recent years. The traditional unsupervised method mainly comprises an image difference method, a ratio method, a principal component analysis method, a change vector analysis method and various statistical modeling methods. These methods are typically based on pixel-level or manually designed features, and are susceptible to interference from factors such as illumination variations, atmospheric condition differences, seasonal variations, and sensor noise, and are not robust. With the development of deep learning technology, a part of unsupervised methods begin to introduce a self-encoder, generate a countermeasure network and other deep architecture, and realize change detection through a style alignment or reconstruction mechanism of cross-phase images. Such methods map images of one phase to the style space of another phase, regarding regions that are difficult to align or reconstruct as varying regions, thereby improving detection performance to some extent. However, the existing unsupervised change detection method based on style alignment still has obvious defects that firstly, the optimization process is usually dominated by style loss, the model tends to reduce overall loss by excluding areas with small contribution to loss optimization (usually changing areas) to form a so-called optimization shortcut problem, secondly, due to overfocusing on style differences, the semantic change of a geospatial structure is ignored, partial unchanged but similar-style areas are misjudged as changing areas, and some truly changed but similar-style areas are missed, and furthermore, the existing method has limited capability in fine characterization of changing boundaries, and the generated change graph often has the problems of boundary blurring, internal discontinuity or more noise, so that the requirement of high-precision change analysis is difficult to meet. In addition, remote sensing image change detection faces several unique challenges including (1) significant illumination, atmospheric conditions and seasonal differences among two-phase images, (2) the ground feature change generally presents complicated morphology and irregular boundaries in space, (3) the change area generally occupies smaller area in the whole image and has serious category imbalance problems, and (4) the lack of clear supervision signal guidance model in an unsupervised scene focuses on real change semantics instead of surface style differences. These challenges make it particularly difficult to develop an unsupervised method that can compromise change detection accuracy, boundary accuracy, and training stability. Therefore, a new unsupervised change detection method is needed, which can effectively avoid the problem of optimizing shortcuts without relying on m