Hyperspectral image change detection method based on scene-aware parameter injection
By freezing the base model and constructing a style condition parameter injection module, screening sensitive layers and performing lightweight parameter adjustment, the problem of insufficient adaptability of hyperspectral change detection methods under new observation conditions is solved, and efficient and accurate change detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2026-06-16
- Publication Date
- 2026-07-31
AI Technical Summary
Existing hyperspectral change detection methods are not adaptable enough to new observation conditions, leading to false detections and missed detections. Furthermore, full retraining is computationally expensive and lacks adaptive adjustment for imaging style shifts.
The main parameters of the basic change detection model are frozen, a style condition parameter injection module is constructed, and scene-specific parameters are generated through the sensitivity parameter selection and parameter space recalibration module, with only a small number of style-sensitive layers being lightly adjusted.
It enables rapid adaptation under different imaging conditions, reduces parameter update volume and training time, and improves change detection accuracy and adaptation efficiency, making it suitable for scenarios such as urban expansion monitoring, crop growth analysis, and disaster assessment.
Smart Images
Figure CN122493070A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of hyperspectral image processing technology, specifically relating to a method for rapid change detection of hyperspectral images based on scene-aware parameter injection. Background Technology
[0002] Hyperspectral image change detection (HCD) is a technique that uses hyperspectral remote sensing images acquired at different times to identify areas of change on the Earth's surface by analyzing differences in their spatial structure and spectral response. Because hyperspectral images contain rich spectral information and can finely characterize the features of ground materials, they have significant application value in fields such as urban monitoring, precision agriculture, and disaster emergency response.
[0003] In recent years, deep learning-based hyperspectral change detection methods have achieved high accuracy. However, existing methods typically rely on training with fixed dual-temporal images. In practical remote sensing monitoring, satellites or airborne platforms continuously acquire new observation images, forming continuous multi-temporal data. When new observation images arrive, they are often affected by factors such as illumination conditions, atmospheric conditions, sensor noise, imaging angle, hardware degradation, and missing bands. This often results in significant non-semantic style shifts in the images, making it easy for the original change detection model to experience a decrease in detection accuracy under new observation conditions. Existing hyperspectral change detection methods usually require model retraining when receiving new observation data, which incurs significant computational and time overhead, making it difficult to meet the rapid response requirements of continuous remote sensing monitoring. Changes in imaging conditions are a significant factor contributing to this requirement.
[0004] A method for detecting cross-domain variations in hyperspectral images, published under license number CN120852985A, is described. This method can alleviate the distribution differences between the source and target domains and improve the model's detection capability in cross-domain scenarios. However, it mainly relies on source-target domain feature learning, pseudo-label generation, and overall model training. It does not generate scene-specific parameters for the imaging style information of newly observed images in continuous observations, nor does it select a small number of key layers for parameter space recalibration based on the network layers' sensitivity to imaging style changes.
[0005] A lightweight, cross-temporal spatial-spectral feature fusion method, system, device, and medium for hyperspectral change detection, disclosed in publication number CN117115675B, can reduce model complexity and improve hyperspectral change detection efficiency to some extent. However, it mainly focuses on the design of a lightweight detection network structure, and still belongs to the scheme of obtaining a fixed change detection model through training. It does not address the problem of rapid adaptation after the arrival of new observation images in continuous observation, nor does it involve style-sensitive layer selection, scene-specific parameter generation, and parameter space recalibration mechanisms after freezing the base model.
[0006] The main problems with existing technologies are: 1) Existing hyperspectral change detection methods are not adaptable to new observation conditions. Directly using the original model can easily lead to false detections and missed detections, while full retraining has high computational costs.
[0007] 2): Existing model update methods usually lack targeted modeling of imaging style shifts, making it difficult to adaptively adjust using the spectral distribution and noise characteristics of new observation images.
[0008] 3) Existing parameter efficient fine-tuning methods (PEFT) mostly adopt fixed module or fixed parameter update strategies, which do not fully consider the sensitivity of different network layers to style changes, and are prone to redundant calculations or insufficient adaptation effects.
[0009] Therefore, it is necessary to propose a new rapid adaptive method for hyperspectral image change detection. While maintaining the main structure and change recognition ability of the original detection model, a small number of key parameters are adaptively adjusted according to the style information of the new observed image, thereby reducing training costs and improving the change detection accuracy under complex observation conditions. Summary of the Invention
[0010] To overcome the shortcomings of the existing technology, the present invention aims to provide a method for rapid change detection of hyperspectral images based on scene-aware parameter injection. This method generates scene-specific parameters based on the imaging style information of newly observed hyperspectral images after freezing the main parameters of the basic change detection model, and performs lightweight parameter recalibration on the style-sensitive layer, thereby reducing the amount of parameter updates while improving the accuracy and adaptability of change detection under different observation conditions.
[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for detecting rapid changes in hyperspectral images based on scene-aware parameter injection includes the following steps; Step 1: Acquire the initial dual-temporal hyperspectral images, pre-train the basic change detection model and freeze its main parameters to serve as the parameter basis for subsequent style-sensitive layer screening and parameter space recalibration; Step 2: Construct the Style Conditional Parameter Injection Module (SPI), which includes a Sensitivity Parameter Selection Module (SPS) and a Parameter Space Recalibration Module (PSR). Step 3: Acquire the new observation hyperspectral image and use it as the input to the Style Condition Parameters (SPI) module to obtain the style feature information of the current observation scene; Step 4: Based on the Sensitivity Parameter Selection Module (SPS), determine the style-sensitive layer to be injected from the basic change detection model pre-trained and frozen in Step 1. Step 5: Extract style cue features based on the newly observed hyperspectral image and generate scene-specific parameters corresponding to the style-sensitive layer to be injected; Step 6: Lightweight parameter adjustment of the style-sensitive layer to be injected obtained in Step 4 is performed through the parameter space recalibration module PSR to obtain the recalibrated basic change detection model. Step 7: Use the recalibrated basic change detection model to generate the corresponding change detection results for the new observation.
[0012] Step 1 specifically involves: Acquire reference phase hyperspectral image and the initial observation phase hyperspectral image ,in H and W represent the spatial height and width of the image, respectively. Indicates the number of spectral bands, and Input the basic change detection model to obtain the initial change prediction map: in, This represents the basic change detection model. The model parameters are represented by the change labels used to pre-train the basic change detection model based on the change labels. The change labels are pixel-level binary change labels used to mark the changed and unchanged pixels between the reference time-phase image and the initial observation time-phase image. The main parameters are obtained after training. ,parameter This represents the original basic parameters obtained by the basic change detection model after pre-training on the initial two-phase images T1 and T2; In this model, each layer of the basic change detection model is called a candidate network layer.
[0013] The main parameters are frozen to preserve the ability to identify semantic changes in ground features.
[0014] Step 2 specifically involves: A Style Conditional Parameter Injection (SPI) module is constructed to perform lightweight adaptive adjustment of the frozen basic change detection model when new observation images arrive. The SPI module includes a style feature extractor, a sensitivity parameter selection module (SPS), and a parameter space recalibration module (PSR). The style feature extractor consists of a 1×1 convolution, a non-linear activation function, and a global average pooling operation, used to extract its style cue features; The Sensitivity Parameter Selection (SPS) module is used to select network layers that are sensitive to changes in imaging style from the basic change detection model; the Parameter Space Recalibration (PSR) module is used to generate scene-specific parameters based on the spectral style information of the newly observed image and to adjust the parameters of the selected style-sensitive layers.
[0015] SPS stands for Sensitivity Parameter Selection Strategy, which is used to select network layers from the basic change detection model that are sensitive to changes in imaging style. PSR is also a strategy that uses a parameter space recalibration mechanism to perform lightweight parameter adjustment of the style-sensitive layer to be injected through matrix decomposition and merging operations.
[0016] Step 3 specifically involves: Acquire new hyperspectral images obtained during continuous observation. The newly observed hyperspectral image Hyperspectral image with reference time phase To construct a new image to be detected: = in, This represents the image pair to be detected corresponding to the new observation, since The image may be affected by factors such as lighting conditions, atmospheric conditions, sensor noise, imaging angle, hardware degradation, and missing bands, resulting in an imaging style shift compared to the initial training images. The Input Style Conditional Parameter Injection (SPI) module is used for subsequent style cue feature extraction and parameter recalibration.
[0017] The style feature extractor of SPI generates style feature information of the current new observation scene through 1×1 convolution, non-linear activation function and global average pooling operation, so as to facilitate subsequent parameter recalibration.
[0018] Step 4 specifically involves: The style-sensitive layer to be injected is determined based on the sensitivity parameter selection module (SPS). This is first done using the initial observation phase hyperspectral image. Perform spatial-spectral perturbation to generate perturbation images The perturbation methods include one or more of Gaussian noise addition, spectral band perturbation, stripe loss, spatial jitter, and local pixel replacement; Next, temporary parameter injectors are introduced into each candidate network layer of the basic change detection model, and perturbation images are used. right Short-range optimization of the temporary parameter injector; for the first There are 1 network layers, and their original parameters are . The temporary injection parameters are The style sensitivity of this network layer is then measured by the angle offset in the parameter space: in, Indicates the first Style sensitivity of each network layer This indicates a parameter flattening operation. Represents the vector dot product. The larger the value, the more significant the parameter orientation shift of the layer under perturbation conditions, and the more sensitive it is to changes in imaging style; Finally, the sensitivity of all candidate layers is ranked, and the top-ranked layers are selected. The set of style-sensitive layers to be injected consists of several style-sensitive layers: in, This represents a set of style-sensitive layers. Indicates the number of network layers selected. This indicates the k-th selected network layer.
[0019] The specific network structure of the basic change detection model is not limited and can be any existing hyperspectral image change detection model. The Style Conditional Parameter Injection (SPI) module described in this invention is a plug-and-play module that can be inserted into the basic change detection model. It recalibrates the parameters of a small number of style-sensitive layers while freezing the original basic model parameters, thereby improving the model's change detection accuracy and adaptation efficiency under new observation conditions.
[0020] Step 5 specifically involves: Based on the extraction of style cue features from newly observed hyperspectral images and the generation of scene-specific parameters, a lightweight style encoder is first constructed to process the newly observed hyperspectral images. The input style encoder extracts its style cue features through 1×1 convolution, non-linear activation functions, and global average pooling operations: in, Indicates style encoder, Indicates style encoder parameters, Style hints and features indicating newly observed images; Then, hierarchical embedding features are constructed for each style-sensitive layer to be injected. The style cue features and hierarchical embedding features of the newly observed image are input together into a parameter-sharing synthesizer to generate the first... Scene-specific parameters for a style-sensitive layer to be injected: in, and This represents the synthesized matrix of learnable parameters shared across layers. This represents element-wise multiplication. Indicates the first The scene-specific parameters corresponding to the style-sensitive layer to be injected.
[0021] Step 6 specifically involves: A lightweight parameter adjustment is performed on the style-sensitive layers to be injected using a parameter space recalibration mechanism. The first in Each network layer, first its original parameters Perform low-rank decomposition: And extract the low-rank left and right basis: in, This represents the rank of a low-rank decomposition. and They represent the first The low-rank left basis and low-rank right basis of each network layer; Subsequently, the first generated in step 5 Scene-specific parameters for a style-sensitive layer to be injected. Divided into two modulation sections and Then, it is incorporated into the low-rank parameter space and the original parameters are recalibrated: in, Indicates the number after recalibration Each network layer parameter, and This represents the parameter adjustment intensity coefficient, used to control the parameter correction ratio, for parameters not selected into the set. The network layers have their parameters frozen.
[0022] Step 7 specifically involves: Reference temporal hyperspectral image With new observations of hyperspectral images Input the change detection model after parameter recalibration to obtain the corresponding change detection prediction map at the current observation time: in, This represents the set of model parameters after style condition parameter injection. This represents the change detection and prediction graph corresponding to the new observation; Finally, threshold segmentation or category discrimination is performed on the predicted map to obtain a binarized change detection map: in, Indicates the first The final detection result for each pixel is 1, indicating a changed pixel and 0, indicating an unchanged pixel. This represents the detection threshold for a continuous observation sequence. Repeat steps 3 to 7 to complete the detection of changes in multi-temporal hyperspectral images.
[0023] The method is applicable to scenarios with high requirements for continuous observation efficiency, detection accuracy, and model deployment cost, specifically urban expansion monitoring, crop growth analysis, and disaster assessment.
[0024] The beneficial effects of this invention are: (1) This invention freezes the main parameters of the basic change detection model and generates scene-specific parameters by using the spectral style information of newly observed hyperspectral images, thereby achieving rapid adaptation under different imaging conditions and reducing the amount of parameter updates and training time required for full retraining.
[0025] (2) This invention uses a sensitivity parameter selection strategy to screen key network layers that are sensitive to changes in imaging style, avoiding the need to update all network layers uniformly, thus reducing the problem of insufficient adaptation caused by redundant calculations and blindly selecting update layers.
[0026] (3) This invention uses a parameter space recalibration mechanism to perform lightweight adjustments on a small number of style-sensitive layers in the low-rank parameter space, which can alleviate the interference of non-semantic factors such as illumination changes, atmospheric disturbances, sensor noise, and band degradation on the change detection results. This invention uses the parameter space recalibration module PSR to convert style cue features extracted from newly observed hyperspectral images into scene-specific parameters, and then performs parameter recalibration on a small number of style-sensitive layers in the low-rank parameter space. The principle is that PSR does not perform a full parameter update of the basic change detection model, but rather makes small corrections to the original parameters based on the low-rank parameter basis, enabling the model parameters to adapt to the imaging style changes of the currently observed image. Therefore, this invention can alleviate the interference of non-semantic factors such as illumination changes, atmospheric disturbances, sensor noise, and band degradation on the change detection results while maintaining the stability of the main parameters of the original model.
[0027] (4) This invention has good model compatibility and scalability, and can be applied to different change detection models such as convolutional neural networks, Transformer networks, state-space models, and hyperspectral basic models. It is suitable for continuous remote sensing monitoring and multi-temporal hyperspectral change detection tasks. The Style Conditional Parameter Injection (SPI) module in this invention has plug-and-play characteristics and does not limit the specific network structure of the basic change detection model. It filters the style-sensitive layers in different basic models through SPS and recalibrates the parameters of the style-sensitive layers through PSR. Therefore, it is compatible with different change detection models such as convolutional neural networks, Transformer networks, state-space models, and hyperspectral basic models. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0029] Figure 2 This is a schematic diagram of the overall structure of the present invention.
[0030] Figure 3 This is a schematic diagram of the parameter space recalibration process in this invention. Detailed Implementation
[0031] The present invention will now be described in further detail with reference to the accompanying drawings.
[0032] To achieve the above objectives, this invention provides a method for rapid change detection in hyperspectral images based on scene-aware parameter injection, such as... Figure 1 As shown, the technical solution adopted includes the following steps: Step 1: Acquire initial dual-temporal hyperspectral images, pre-train the basic change detection model, and freeze its parameters: This step is used to obtain a basic change detection model with initial ground feature change recognition capabilities, providing a stable model foundation for rapid adaptation in subsequent continuous observation scenarios.
[0033] First, acquire the reference temporal hyperspectral image. and the initial observation phase hyperspectral image The two hyperspectral images were preprocessed, including normalization, band alignment, and spatial registration, to ensure they had consistent spatial dimensions and spectral dimensions. H, W, and C represent the spatial height, spatial width, and number of spectral bands of the hyperspectral image, respectively.
[0034] Then, Input the basic change detection model to obtain the initial two-phase change prediction map: in, This represents the basic change detection model. Indicates model parameters, This represents the initial two-phase change prediction diagram.
[0035] The basic change detection model is pre-trained using change labels. After the model training is complete, the pre-trained parameters are obtained. The main parameters of the basic change detection model are frozen so that it can maintain its original semantic recognition ability of ground feature changes during subsequent continuous observations, thus avoiding repeated training of the entire model.
[0036] Step 2: Construct the Style Conditional Parameter Injection (SPI) module, which includes a Sensitivity Parameter Selection (SPS) module and a Parameter Space Recalibration (PSR) module. This step involves constructing a lightweight module that can dynamically adjust the parameters of a fundamental change detection model based on the imaging style of newly observed hyperspectral images. Without altering the core structure of the fundamental change detection model, this module generates a small number of parameter corrections, enabling the model to quickly adapt to different observation conditions.
[0037] The Style Conditional Parameter Injection (SPI) module includes a Sensitivity Parameter Selection (SPS) module and a Parameter Space Recalibration (PSR) module. The SPS module selects network layers from the basic change detection model that are highly sensitive to changes in imaging style. The PSR module generates scene-specific parameters based on the spectral style information of newly observed hyperspectral images and performs lightweight parameter adjustments on the selected style-sensitive layers.
[0038] This module does not update all parameters of the basic change detection model, but only generates parameter corrections for a small number of style-sensitive layers, thereby reducing training overhead and lowering the deployment cost of the model in continuous observation scenarios.
[0039] Step 3: Acquire the newly observed hyperspectral image and inject it as input to the Style Condition Parameters (SPI) module: This step is used to acquire new observation images that arrive during continuous observation and to determine any possible imaging style shifts between them and the initial observation images.
[0040] In actual continuous remote sensing monitoring, satellites or airborne platforms continuously acquire new observation images, forming multi-temporal hyperspectral sequences. Let the currently arriving new hyperspectral image be... Then it is compared with the reference time-phase hyperspectral image. Constructing a new image pair to be detected .
[0041] Due to newly observed hyperspectral images It may be affected by factors such as illumination conditions, atmospheric conditions, sensor noise, imaging angle, hardware degradation, and missing bands, resulting in a difference between its image and the initial hyperspectral image at the time of observation. There is often an image style shift between them. If the frozen base change detection model is used directly for detection, problems such as false positives and false negatives are likely to occur.
[0042] Therefore, this step will involve newly observed hyperspectral images. The input style condition parameter injection module is used to extract style cue features of the current observation scene and provide input basis for subsequent scene-specific parameter generation and parameter space recalibration.
[0043] Step 4: Based on the sensitivity parameter selection strategy, determine the style-sensitive layer to be injected from the basic change detection model. This step is used to select network layers that are more sensitive to imaging style shifts from the frozen base change detection model, avoiding redundant calculations caused by uniformly updating all network layers, and reducing the risk of insufficient adaptation due to blindly selecting update layers.
[0044] First, the initial observation time-phase hyperspectral image Perform spatial-spectral perturbations to generate perturbation images simulating new observation conditions. The perturbation methods include one or more of the following: Gaussian noise addition, spectral band perturbation, stripe loss, spatial jitter, and local pixel replacement, used to simulate imaging condition variations that may occur in actual observations.
[0045] Next, temporary parameter injectors are introduced into each candidate network layer of the basic change detection model, and perturbation images are used to... Short-range optimization is performed on the temporary parameter injector. For the first... A network layer, assuming its original... The parameters are The temporary injection parameters are The sensitivity of the layer to changes in imaging style is then measured by the angular offset between the original parameters and the injected parameters. in, Indicates the first Style sensitivity of each network layer This indicates a parameter flattening operation. This represents the vector dot product. The larger the value, the more significant the parameter orientation shift of the layer under perturbation conditions, and the more sensitive it is to changes in imaging style.
[0046] Finally, the sensitivity of all candidate layers is ranked, and the top-ranked layers are selected. The set of style-sensitive layers to be injected consists of several style-sensitive layers: Where S represents the set of style-sensitive layers to be injected, and K represents the number of network layers selected. This indicates the k-th selected network layer. Using this strategy, the model only performs style adaptation on a small number of style-sensitive layers, reducing the scale of parameter updates while maintaining the original change recognition capability of the base model.
[0047] Step 5: Extract style cue features based on the newly observed hyperspectral image and generate scene-specific parameters corresponding to the style-sensitive layer to be injected: This step is used to generate parameter adjustment information suitable for the current observation scenario by utilizing the spectral distribution, noise status, and imaging condition information contained in the newly observed hyperspectral image.
[0048] First, a lightweight style encoder is constructed to process the newly observed hyperspectral images. The input style encoder extracts style cue features from the current image through 1×1 convolution, non-linear activation functions, and global average pooling operations: in, Indicates style encoder, Indicates style encoder parameters, This indicates style cue features for newly observed images. These features characterize non-semantic factors such as illumination conditions, noise distribution, spectral response differences, and sensor degradation in the current image.
[0049] Subsequently, hierarchical embedding features are constructed for each style-sensitive layer to be injected. This is used to represent the parameter space attributes of different network layers. Style cue features... With hierarchical embedding features Input parameters are shared by the synthesizer to generate the first... Scene-specific parameters for a style-sensitive layer to be injected: in, and This represents the parameter composition matrix shared across layers. This represents element-wise multiplication. Indicates the first The scene-specific parameters corresponding to the style-sensitive layer to be injected.
[0050] By employing a parameter-sharing synthesizer, this invention avoids building an independent parameter generation network for each network layer, further reducing the number of training parameters and computational complexity of the style conditional parameter injection module.
[0051] Step 6: Lightweight parameter adjustment of the style-sensitive layer using a parameter space recalibration mechanism. This step aims to inject the scene-specific parameters generated in step 5 into the style-sensitive layer parameters of the basic change detection model, thereby enabling rapid adaptation to the current observation conditions.
[0052] For the first style-sensitive layer in the set S to be injected... Each network layer first freezes its original parameters. Perform low-rank decomposition to obtain a low-dimensional basis representation in the parameter space: And extract the low-rank left and right basis: in, This represents the rank of a low-rank decomposition. and They represent the first The low-rank parameter basis of each network layer.
[0053] Then, the scene-specific parameters Divided into two modulation sections and Then, it is incorporated into the low-rank parameter space and the original parameters are recalibrated: in, Indicates the number after recalibration Each network layer parameter, and Indicates scene-specific parameters The two obtained modulation vectors or modulation matrices are β, which represents the parameter adjustment intensity coefficient and is used to control the parameter correction ratio.
[0054] For network layers not selected into the set S of style-sensitive layers to be injected, their parameters remain frozen. In this way, the present invention introduces lightweight parameter corrections only on a small number of style-sensitive layers, thereby achieving rapid adaptation to new observational imaging styles while maintaining the stability of the basic detection model structure.
[0055] Step 7: Use the recalibrated change detection model to generate the corresponding change detection results for the new observations: This step is used to perform change detection on the reference temporal image and the newly observed image using the change detection model after style condition parameter injection, and output the corresponding change area at the time of the new observation.
[0056] Specifically, reference time-phase hyperspectral images will be used. With new observations of hyperspectral images Input the recalibrated change detection model to obtain the change prediction map for the current observation phase: in, This represents the set of model parameters after style condition parameter injection. This represents the change detection and prediction map corresponding to the new observation.
[0057] Subsequently, threshold segmentation or category discrimination is performed on the change prediction map to obtain the final binarized change detection result: in, Indicates the first The final detection result for each pixel is 1, indicating a changed pixel and 0, indicating an unchanged pixel. This represents the detection threshold. For continuous observation sequences... Repeat steps 3 to 7 to complete the detection of changes in multi-temporal hyperspectral images.
[0058] The seven steps described above are interconnected as follows: First, step 1 pre-trains the basic change detection model using initial dual-temporal hyperspectral images and freezes its main parameters to provide stable semantic recognition capabilities for ground cover changes in subsequent continuous observations. Next, step 2 constructs a style condition parameter injection module to provide a lightweight parameter adjustment mechanism for the model to quickly adapt to new observation conditions. Then, step 3 acquires newly observed hyperspectral images and uses them as a source of style information. Step 4 determines the key layers sensitive to imaging style changes using a sensitivity parameter selection strategy. Step 5 extracts style cue features from the newly observed images and generates scene-specific parameters. Step 6 injects the scene-specific parameters into the style-sensitive layer and recalibrates the basic model in parameter space. Finally, step 7 uses the recalibrated change detection model to output the change detection results for the new observation phase. Through this process, the present invention can achieve rapid, efficient, and high-precision change detection of continuously observed hyperspectral images while maintaining the original model's main structure and change recognition capabilities.
[0059] Experimental conditions: The hardware platform used in the experiment consisted of a single NVIDIA GeForce RTX 5090 GPU with 32GB of video memory. The experimental program was implemented in Python and based on the PyTorch deep learning framework.
[0060] Experimental content and analysis: To verify the effectiveness of this invention in hyperspectral change detection tasks, comparative experiments were conducted on three hyperspectral change detection datasets: BayArea, Farmland, and Santa Barbara, using existing methods such as MSDFFN, LSU-SADM, D2AGCN, ML-EDAN, AIWSEN, BTCDNet, DIEFEN, and HYPERSIGMA. This invention, as a pluggable, fast-adaptive module, was validated using DIEFEN and HYPERSIGMA as representative base detection models to examine its adaptation to conventional change detection models and hyperspectral base models, respectively. The experimental results are shown in Table 1. The Kappa coefficient measures the consistency between the change detection results and the true labels; a value closer to 1 indicates better detection performance. Param (%) represents the proportion of parameters involved in model training or updating; a smaller value indicates lower model update overhead.
[0061] Table 1 As shown in Table 1, under the frozen condition, the overall detection performance of each comparative method is low because no adaptive parameter updates are performed. For example, BTCDNet's Kappa is only 0.6815 on the BayArea dataset, and DIEFEN's Kappa is only 0.6189 on the Santa Barbara dataset, indicating that the directly frozen model is difficult to fully adapt to the differences in spectral distribution and spatial variation characteristics in different hyperspectral scenes.
[0062] Under retraining conditions, the Kappa values of all methods improved compared to the frozen conditions. For example, HYPERSIGMA achieved Kappa values of 0.9608 and 0.9711 on the BayArea and Santa Barbara datasets, respectively, demonstrating good detection performance. However, these methods all have a Param(%) of 100, requiring updates to all network parameters, resulting in high training costs and computational overhead, which is not conducive to rapid model transfer and deployment.
[0063] In contrast, this invention achieves high change detection accuracy with only a very small number of parameter updates. Using DIEFEN as the base model, the Kappa values of this invention on the BayArea, Farmland, and Santa Barbara datasets reach 0.9268, 0.8914, and 0.9637, respectively, while the corresponding Param (%) are only 0.8767, 0.7260, and 0.8767. Using HYPERSIGMA as the base model, the Kappa values of this invention on the three datasets reach 0.9511, 0.9065, and 0.9668, respectively, while the Param (%) are only 0.0034, 0.0033, and 0.0034. Therefore, this invention can achieve detection performance close to that of full-parameter retraining with only a very small number of parameter updates, and even reaches or exceeds the retraining results on some datasets.
[0064] As shown in Figure 2, its workflow is as follows: First, the basic change detection model is pre-trained using initial dual-temporal hyperspectral images and its main parameters are frozen to construct a frozen basic detector. Secondly, style sensitivity analysis of candidate network layers is performed through the Sensitivity Parameter Selection (SPS) module to select a set of style-sensitive layers to be injected. Next, when a new observation image is reached during continuous observation, it is input into the style encoder to extract style cue features, and the parameter space shared synthesizer is used to generate scene-specific parameters corresponding to the style-sensitive layer to be injected. Finally, the scene-specific parameters are injected into the selected style-sensitive layer through the parameter space recalibration module (PSR) to complete the parameter recalibration and output the change detection results of the new temporal phase.
[0065] As shown in Figure 3, after a newly observed hyperspectral image is input into the style encoder, it sequentially undergoes Flatten, global average pooling, nonlinear activation function, and convolution operations to extract style cue features representing the current imaging conditions. Subsequently, these features, along with hierarchical embedded features, are input into a parameter space shared synthesizer to map and generate scene-specific parameters for the corresponding style-sensitive layer. Finally, the scene-specific parameters are divided into two modulation parts and combined with the low-rank left and right basis functions of the original frozen parameters in the low-rank parameter space to generate recalibrated network parameters. This process achieves lightweight parameter adjustment for new observation imaging styles without updating all model parameters. This invention is applicable to scenarios with high requirements for continuous observation efficiency, detection accuracy, and model deployment costs, such as urban expansion monitoring, crop growth analysis, and disaster assessment. For example, in continuous urban expansion monitoring, a basic change detection model is first pre-trained using early dual-temporal hyperspectral images, and its main parameters are frozen. When subsequent observation images for a new quarter or month are acquired, they are input into the Style Conditional Parameter Injection (SPI) module. The sensitivity layer to be injected is determined through Sensitivity Parameter Selection (SPS), and lightweight parameter adjustment is performed using Parameter Space Recalibration (PSR). Finally, by inputting the reference image and the newly observed image into the recalibrated model, the detection results of new buildings, road expansion, or land cover changes can be output efficiently and accurately.
[0066] In summary, this invention demonstrates strong detection capabilities and good parameter efficiency on three hyperspectral change detection datasets. Compared to the frozen model, this invention significantly improves change detection accuracy; compared to the full-parameter retraining method, this invention maintains a high Kappa coefficient while significantly reducing the parameter update ratio. Experimental results show that this invention can effectively improve the model's transfer adaptability in different hyperspectral scenarios, balancing detection accuracy and training efficiency, and has high practical application value.
Claims
1. A method for detecting rapid changes in hyperspectral images based on scene-aware parameter injection, characterized in that, Includes the following steps; Step 1: Acquire the initial dual-temporal hyperspectral images, pre-train the basic change detection model and freeze its main parameters to serve as the parameter basis for subsequent style-sensitive layer screening and parameter space recalibration; Step 2: Construct the Style Conditional Parameter Injection Module (SPI), which includes a Sensitivity Parameter Selection Module (SPS) and a Parameter Space Recalibration Module (PSR). Step 3: Acquire the new observation hyperspectral image and inject it as the input of the style condition parameter module SPI to obtain the style feature information of the current observation scene; Step 4: Based on the Sensitivity Parameter Selection Module (SPS), determine the style-sensitive layer to be injected from the basic change detection model pre-trained and frozen in Step 1. Step 5: Extract style cue features based on the newly observed hyperspectral image and generate scene-specific parameters corresponding to the style-sensitive layer to be injected; Step 6: Lightweight parameter adjustment of the style-sensitive layer to be injected obtained in Step 4 is performed through the parameter space recalibration module PSR to obtain the recalibrated basic change detection model. Step 7: Use the recalibrated basic change detection model to generate the corresponding change detection results for the new observation.
2. The method for rapid change detection of hyperspectral images based on scene-aware parameter injection according to claim 1, characterized in that, Step 1 specifically involves: Acquire reference phase hyperspectral image and the initial observation phase hyperspectral image ,in H and W represent the spatial height and width of the image, respectively. Indicates the number of spectral bands, and Input the basic change detection model to obtain the initial change prediction map: in, This represents the basic change detection model. The model parameters are represented by the change labels used to pre-train the basic change detection model based on the change labels. The change labels are pixel-level binary change labels used to mark the changed and unchanged pixels between the reference time-phase image and the initial observation time-phase image. The main parameters are obtained after training. And freeze the main parameters to preserve the ability to identify semantic changes in ground features; Main parameters This represents the original basic parameters obtained by the basic change detection model after pre-training on the initial two-phase images T1 and T2; In this model, each layer of the basic change detection model is called a candidate network layer.
3. The method for detecting rapid changes in hyperspectral images based on scene-aware parameter injection according to claim 2, characterized in that, Step 2 specifically involves: A Style Conditional Parameter Injection (SPI) module is constructed to perform lightweight adaptive adjustment of the frozen baseline change detection model when new observation images arrive. The SPI module includes a style feature extractor, a sensitivity parameter selection module (SPS), and a parameter space recalibration module (PSR). The style feature extractor consists of a 1×1 convolution, a non-linear activation function, and a global average pooling operation, used to extract its style cue features; Among them, the Sensitivity Parameter Selection (SPS) module is used to select network layers that are sensitive to changes in imaging style from the basic change detection model; the Parameter Space Recalibration (PSR) module is used to generate scene-specific parameters based on the spectral style information of the newly observed image and to adjust the parameters of the selected style-sensitive layers. SPS stands for Sensitivity Parameter Selection Strategy, which is used to select network layers from the basic change detection model that are sensitive to changes in imaging style. PSR is also a strategy that uses a parameter space recalibration mechanism to perform lightweight parameter adjustment of the style-sensitive layer to be injected through matrix decomposition and merging operations.
4. The method for detecting rapid changes in hyperspectral images based on scene-aware parameter injection according to claim 3, characterized in that, Step 3 specifically involves: Acquire new hyperspectral images obtained during continuous observation. The newly observed hyperspectral image Hyperspectral image with reference time phase To construct a new image to be detected: = in, This represents the pair of images to be detected corresponding to the new observation. The input style condition parameter injection module (SPI) generates style feature information of the current new observation scene through the style feature extractor of SPI, which is generated by 1×1 convolution, non-linear activation function and global average pooling operation, so as to facilitate subsequent parameter recalibration.
5. The method for rapid change detection of hyperspectral images based on scene-aware parameter injection according to claim 4, characterized in that, Step 4 specifically involves: First, the initial observation time-phase hyperspectral image was analyzed. Perform spatial-spectral perturbation to generate perturbation images. The perturbation method is one or more of Gaussian noise addition, spectral band perturbation, band loss, spatial jitter, and local pixel replacement; then, temporary parameter injectors are introduced into each candidate network layer of the basic change detection model, and the perturbation image is used. right Short-range optimization of the temporary parameter injector; for the first There are 1 network layers, and their original parameters are . The temporary injection parameters are The style sensitivity of this network layer is then measured by the angle offset in the parameter space: in, Indicates the first Style sensitivity of each network layer This indicates a parameter flattening operation. Represents the vector dot product. The larger the value, the more significant the parameter orientation shift of the layer under perturbation conditions, and the more sensitive it is to changes in imaging style; Finally, the sensitivity of all candidate layers is ranked, and the top-ranked layers are selected. The set of style-sensitive layers to be injected consists of several style-sensitive layers: in, This represents a set of style-sensitive layers. Indicates the number of network layers selected. This indicates the k-th selected network layer.
6. The method for detecting rapid changes in hyperspectral images based on scene-aware parameter injection according to claim 5, characterized in that, The basic change detection model is any existing hyperspectral image change detection model, and the style condition parameter injection module (SPI) is a plug-and-play module.
7. The method for rapid change detection of hyperspectral images based on scene-aware parameter injection according to claim 5, characterized in that, Step 5 specifically involves: First, a lightweight style encoder is constructed to process the newly observed hyperspectral images. The input style encoder extracts its style cue features through 1×1 convolution, non-linear activation functions, and global average pooling operations: in, Indicates style encoder, Indicates style encoder parameters, Style hints and features indicating newly observed images; Then, hierarchical embedding features are constructed for each style-sensitive layer to be injected. The style cue features and hierarchical embedding features of the newly observed image are input together into a parameter-sharing synthesizer to generate the first... Scene-specific parameters for a style-sensitive layer to be injected: in, and This represents the synthesized matrix of learnable parameters shared across layers. This represents element-wise multiplication. Indicates the first The scene-specific parameters corresponding to the style-sensitive layer to be injected.
8. The method for rapid change detection of hyperspectral images based on scene-aware parameter injection according to claim 7, characterized in that, Step 6 specifically involves: A lightweight parameter adjustment is performed on the style-sensitive layers to be injected using a parameter space recalibration mechanism. The first in Each network layer, first its original parameters Perform low-rank decomposition: And extract the low-rank left and right basis: in, This represents the rank of a low-rank decomposition. and They represent the first The low-rank left basis and low-rank right basis of each network layer; Subsequently, the first generated in step 5 Scene-specific parameters for a style-sensitive layer to be injected. Divided into two modulation sections and Then, it is incorporated into the low-rank parameter space and the original parameters are recalibrated: in, Indicates the number after recalibration Each network layer parameter, and This represents the parameter adjustment intensity coefficient, used to control the parameter correction ratio, for parameters not selected into the set. The network layers have their parameters frozen.
9. The method for rapid change detection of hyperspectral images based on scene-aware parameter injection according to claim 8, characterized in that, Step 7 specifically involves: Reference temporal hyperspectral image With new observations of hyperspectral images Input the change detection model after parameter recalibration to obtain the corresponding change detection prediction map at the current observation time: in, This represents the set of model parameters after style condition parameter injection. This represents the change detection and prediction graph corresponding to the new observation; Finally, threshold segmentation or category discrimination is performed on the predicted map to obtain a binarized change detection map: in, Indicates the first The final detection result for each pixel is 1, indicating a changed pixel and 0, indicating an unchanged pixel. This represents the detection threshold for a continuous observation sequence. Repeat steps 3 to 7 to complete the detection of changes in multi-temporal hyperspectral images.
10. The application of the method according to any one of claims 1-9, characterized in that, The method is applicable to scenarios with high requirements for continuous observation efficiency, detection accuracy, and model deployment cost, specifically urban expansion monitoring, crop growth analysis, and disaster assessment.