Remote sensing image change detection method and system based on physical imaging parameter fusion

Abstract

This invention discloses a method and system for detecting changes in remote sensing images based on physical imaging parameter fusion. The method includes: acquiring a first-temporal UAV remote sensing image and a second-temporal UAV remote sensing image; performing metadata parsing and encoding to obtain metadata embedding features; performing feature extraction processing on the first-temporal and second-temporal UAV remote sensing images based on an image feature extraction encoder to obtain multi-scale dual-temporal visual features; applying conditional modulation to the multi-scale dual-temporal visual features as conditional input, and performing decoding output processing to obtain a UAV remote sensing change map. This invention can achieve active compensation for optical distortion and scale inconsistency, improving the detection accuracy and robustness of UAV remote sensing changes. As a method and system for detecting changes in remote sensing images based on physical imaging parameter fusion, this invention can be widely applied in the field of computer vision technology.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for detecting changes in remote sensing images based on the fusion of physical imaging parameters. Background Technology

[0002] Remote sensing change detection is a technique that automatically compares and analyzes images of the same geographic area acquired at different times to identify meaningful changes in ground features. In recent years, deep learning methods, represented by Siamese convolutional neural networks (such as SNUNet and Fang), and Transformer-based methods (such as BIT and Chen) have made significant progress on publicly available change detection benchmark datasets. However, these methods all share a common drawback: when processing UAV remote sensing images, they fail to fully utilize the EXIF ​​physical metadata in the UAV image files and their associated flight records, and fail to utilize prior physical information directly related to imaging conditions, such as flight altitude and attitude angles.

[0003] Therefore, the relevant technologies have the following problems: First, optical distortion problem. Radial distortion of wide-angle and fisheye lenses causes nonlinear deformation of ground features in the edge areas of the image. Existing pixel domain pre-correction schemes have inherent limitations such as introducing interpolation errors and decoupling from network feature extraction. Second, multi-temporal scale inconsistency problem. Differences in flight altitude cause mismatch in ground sampling distances between two-temporal images. Existing random multi-scale data augmentation only performs statistical compensation and cannot use the actual flight altitude of the test image for targeted compensation. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a method and system for detecting changes in remote sensing images based on the fusion of physical imaging parameters. This method and system can actively compensate for optical distortion and scale inconsistencies, thereby improving the detection accuracy and robustness of changes in UAV remote sensing.

[0005] The first technical solution adopted in this invention is: a remote sensing image change detection method based on physical imaging parameter fusion, comprising the following steps: Acquire UAV remote sensing images from the first and second time phases, perform metadata parsing and encoding processing, and obtain metadata embedding features; Based on the image feature extraction encoder, feature extraction processing is performed on the first phase UAV remote sensing image and the second phase UAV remote sensing image to obtain multi-scale dual-temporal visual features. Metadata embedding features are used as conditional inputs to apply conditional modulation to multi-scale dual-temporal visual features, and then the decoded output is processed to obtain a UAV remote sensing change map.

[0006] Furthermore, the step of acquiring the first-time-phase UAV remote sensing image and the second-time-phase UAV remote sensing image, performing metadata parsing and encoding processing to obtain metadata embedding features specifically includes: The first and second time-phase UAV remote sensing images are acquired, and the metadata of the physical imaging parameters is parsed to obtain the metadata of the remote sensing images. The metadata includes flight altitude, position information, lens focal length, attitude angle, sensor resolution and shooting time information. The metadata of the remote sensing image is normalized to obtain the first single-temporal metadata feature vector and the second single-temporal metadata feature vector; Obtain the dual-phase difference feature vector of the first single-phase metadata feature vector and the second single-phase metadata feature vector, and concatenate them with the first single-phase metadata feature vector and the second single-phase metadata feature vector to construct an extended metadata vector; The extended metadata vector is input into the metadata encoder for encoding to obtain the metadata embedding features.

[0007] Furthermore, the specific expression for calculating the dual-phase difference feature vector is as follows: ; In the above formula, This represents the eigenvectors representing the differences between the two time phases. This represents the first single-phase metadata feature vector. This represents the feature vector of the second single-phase metadata.

[0008] Furthermore, both the first and second single-phase metadata feature vectors include the logarithmically normalized value of flight altitude, the trigonometric function encoded value of attitude angle, the encoded value of position offset, the normalized ground sampling distance value, and the encoded value of time period. The dual-phase difference feature vector includes the dual-phase flight altitude encoding difference, the ground sampling distance encoding difference, and the yaw angle encoding difference.

[0009] Furthermore, the step of performing feature extraction processing on the first-temporal UAV remote sensing image and the second-temporal UAV remote sensing image based on the image feature extraction encoder to obtain multi-scale dual-temporal visual features specifically includes: The first-time-phase UAV remote sensing image and the second-time-phase UAV remote sensing image are input into the image feature extraction encoder, which includes an image feature extraction module, a shared weight VSSM encoder and an MSTFM fusion module. The image feature extraction module based on the image feature extraction encoder performs feature extraction processing on the first phase UAV remote sensing image and the second phase UAV remote sensing image to obtain the features of the first phase UAV remote sensing image and the features of the second phase UAV remote sensing image. Based on the image feature extraction encoder, the shared weight VSSM encoder performs weight-sharing encoding on the features of the first phase UAV remote sensing image and the features of the second phase UAV remote sensing image, to obtain the encoded features of the first phase UAV remote sensing image and the encoded features of the second phase UAV remote sensing image. The MSTFM fusion module based on the image feature extraction encoder performs feature fusion on the encoded first-temporal UAV remote sensing image features and the encoded second-temporal UAV remote sensing image features to obtain multi-scale dual-temporal visual features.

[0010] Furthermore, the step of using metadata embedding features as conditional input to apply conditional modulation to multi-scale dual-temporal visual features and performing decoding output processing to obtain the UAV remote sensing change map specifically includes: Metadata embedding features are used as conditional inputs, and conditional modulation is applied to multi-scale dual-temporal visual features through feature-level affine transformation to obtain conditionally modulated remote sensing image features. The features of the conditionally modulated remote sensing image are sequentially classified and upsampled to obtain a pixel-level change probability map. Thresholding is performed on the pixel-level change probability map to obtain the UAV remote sensing change map.

[0011] Furthermore, the feature-level affine transformation represents the dynamic generation of channel-level scale vectors and channel-level offset vectors from metadata embedded features via linear projection, and the injection of these vectors into the feature stream with zero-initialized residuals, wherein: The expression for the channel-level scale vector is as follows: ; The expression for the channel-level offset vector is as follows: ; The specific expression for the modulation result of the feature-level affine transformation is as follows: ; The specific expression for injecting the feature stream using the zero-initialization residual method is as follows: ; In the above formula, Represents the channel-level scale vector. Represents the channel-level offset vector. Represents metadata embedding features, Indicates the first Phase Decoding features, Indicates a learnable scalar. This indicates channel-by-channel multiplication. This represents the linear projection weight matrix that generates the channel-level scale vector. This represents the linear projection bias that generates the channel-level scale vector. This represents the linear projection weight matrix that generates the channel-level offset vector. This represents the linear projection bias used to generate the channel-level offset vector. Represents the channel-level scale vector. Represents the channel-level offset vector. This indicates the decoding characteristics after modulation.

[0012] Furthermore, it also includes constructing a distortion-aware weighted loss function by combining the weighted binary cross-entropy loss function and the Lovász-Softmax loss function during the model training phase.

[0013] Furthermore, it also includes building a metadata missing fault tolerance mechanism. When some parameters in the metadata are missing, the missing components are filled with the mean of the corresponding parameters in the training set. When all metadata is missing, the metadata embedding is replaced with an all-zero vector.

[0014] The second technical solution adopted in this invention is: a remote sensing image change detection system based on physical imaging parameter fusion, comprising: The first module is used to acquire the first phase UAV remote sensing image and the second phase UAV remote sensing image, perform metadata parsing and encoding processing, and obtain metadata embedding features; The second module is used to perform feature extraction processing on the first phase UAV remote sensing image and the second phase UAV remote sensing image based on the image feature extraction encoder to obtain multi-scale dual-temporal visual features. The third module is used to take the metadata embedding features as conditional input, apply conditional modulation to the multi-scale dual-temporal visual features, and perform decoding output processing to obtain the UAV remote sensing change map.

[0015] The beneficial effects of the method and system of this invention are as follows: This invention acquires a first-phase UAV remote sensing image and a second-phase UAV remote sensing image, performs metadata parsing and encoding processing to obtain metadata embedding features, and systematically integrates EXIF ​​physical metadata into a deep learning change detection framework to achieve active compensation for optical distortion and scale inconsistency; further, based on an image feature extraction encoder, it performs feature extraction processing on the first-phase and second-phase UAV remote sensing images to obtain multi-scale dual-temporal visual features; finally, it uses the metadata embedding features as conditional input to apply conditional modulation to the multi-scale dual-temporal visual features and performs decoding output processing to obtain a UAV remote sensing change map. Feature-level affine transformation applies dynamic compensation independently to each channel, which improves accuracy compared to feature stitching. The EXIF ​​physical parameters are integrated into the change detection network in a feature-level conditional modulation manner to achieve active compensation for optical distortion and scale inconsistency, thereby improving the detection accuracy and robustness of UAV remote sensing changes. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of the remote sensing image change detection method based on physical imaging parameter fusion of the present invention; Figure 2 This is a structural block diagram of the remote sensing image change detection system based on physical imaging parameter fusion of the present invention; Figure 3 This is a schematic diagram of the step framework for remote sensing image change detection provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the EXIF ​​metadata parsing and feature vector construction module provided in a specific embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0018] Reference Figure 1 and Figure 3 This invention provides a method for detecting changes in remote sensing images based on the fusion of physical imaging parameters. The method includes the following steps: S100: Acquire the first-time-phase UAV remote sensing image and the second-time-phase UAV remote sensing image, perform metadata parsing and encoding processing, and obtain metadata embedding features; Specifically, a first-temporal UAV remote sensing image and a second-temporal UAV remote sensing image are acquired, and metadata parsing processing of physical imaging parameters is performed to obtain metadata of the remote sensing images. The metadata includes flight altitude, position information, lens focal length, attitude angle, sensor resolution, and shooting time information. The metadata of the remote sensing images is normalized to obtain a first-temporal metadata feature vector and a second-temporal metadata feature vector. The dual-temporal difference feature vector of the first-temporal metadata feature vector and the second-temporal metadata feature vector is obtained and concatenated with the first-temporal metadata feature vector and the second-temporal metadata feature vector to construct an extended metadata vector. The extended metadata vector is input to the metadata encoder for encoding processing to obtain metadata embedding features.

[0019] In this embodiment, the physical imaging parameters in the UAV image files and their associated flight records are transformed into prior knowledge of the deep learning change detection network, and feature-level conditional modulation is used to actively compensate for differences in imaging conditions.

[0020] First, acquire dual-temporal UAV remote sensing images I1 and I2 of the same geographical area, and parse metadata (such as EXIF ​​format) containing physical imaging parameters from the corresponding image files or their associated metadata files. This metadata includes at least flight altitude H, GPS coordinates, lens focal length f, and sensor resolution. Further, normalize the EXIF ​​metadata to construct single-temporal metadata feature vectors m1 and m2, each containing at least one encoded component of a physical imaging parameter. Calculate the dual-temporal difference feature vector, whose expression is: ; In the above formula, This represents the eigenvectors representing the differences between the two time phases. This represents the first single-phase metadata feature vector. This represents the feature vector of the second single-phase metadata.

[0021] The difference features are concatenated with the feature vectors of each temporal phase to form an extended metadata vector. The difference feature vector explicitly encodes the changes in flight conditions in both temporal phases. The extended metadata vector is then input into the metadata feature encoder to generate a high-dimensional metadata embedding representation. .

[0022] Additionally, it should be noted that the UAV remote sensing images are panoramic images taken with wide-angle or fisheye lenses, which have radial and tangential distortion. The metadata feature vector also includes a distortion intensity estimation component with the normalized radial distance r / rmax from the pixel coordinates to the image principal point as a proxy variable. This component is used to characterize the degree of distortion at each location in the rectangular projection panoramic image and can be obtained without camera calibration parameters.

[0023] In some specific embodiments, such as Figure 4 As shown, the single-temporal metadata feature vector contains the following types of encoded components: log-normalized flight altitude, trigonometric function encoded attitude angle, encoded position offset, normalized ground sampling distance, and encoded time period; the difference feature vector contains the difference in flight altitude encoding between the two temporal phases, the difference in ground sampling distance encoding, and the difference in yaw angle encoding; the single-temporal metadata feature vector and Each has 9 dimensions, the difference feature vector Also 9-dimensional, the extended metadata vector is composed of bi-temporal feature vectors. , With difference feature vector The three components are combined to form a total of 27 dimensions.

[0024] In addition, the metadata feature encoder is a multilayer perceptron, which contains at least two fully connected layers, each followed by a batch normalization layer and a nonlinear activation function; the encoder output dimension is not less than 64 dimensions; the number of encoder parameters does not exceed 0.05M.

[0025] S200: Based on the image feature extraction encoder, feature extraction processing is performed on the first phase UAV remote sensing image and the second phase UAV remote sensing image to obtain multi-scale dual-temporal visual features. Specifically, the first-temporal UAV remote sensing image and the second-temporal UAV remote sensing image are input into an image feature extraction encoder, which includes an image feature extraction module, a shared-weight VSSM encoder, and an MSTFM fusion module. Based on the image feature extraction encoder, the image feature extraction module performs feature extraction processing on the first-temporal UAV remote sensing image and the second-temporal UAV remote sensing image to obtain the first-temporal UAV remote sensing image features and the second-temporal UAV remote sensing image features. Based on the shared-weight VSSM encoder of the image feature extraction encoder, the first-temporal UAV remote sensing image features and the second-temporal UAV remote sensing image features are subjected to weight-sharing encoding processing to obtain the encoded first-temporal UAV remote sensing image features and the encoded second-temporal UAV remote sensing image features. Based on the image feature extraction encoder, the MSTFM fusion module performs feature fusion on the encoded first-temporal UAV remote sensing image features and the encoded second-temporal UAV remote sensing image features to obtain multi-scale dual-temporal visual features.

[0026] In this embodiment, the dual-temporal images I1 and I2 are respectively input into a shared-weight image feature extraction encoder to extract multi-scale dual-temporal visual features.

[0027] S300. Using metadata embedding features as conditional input, conditional modulation is applied to multi-scale dual-temporal visual features, and decoding output processing is performed to obtain the UAV remote sensing change map.

[0028] Specifically, metadata embedding features are used as conditional inputs, and conditional modulation is applied to multi-scale dual-temporal visual features through feature-level affine transformation to obtain conditionally modulated remote sensing image features. The conditionally modulated remote sensing image features are then classified and upsampled sequentially to obtain pixel-level change probability maps. Finally, the pixel-level change probability maps are thresholded to obtain UAV remote sensing change maps.

[0029] In this embodiment, metadata is embedded during at least one feature refinement stage of the change detection decoder. As a conditional input, the visual features are conditionally modulated through a feature-level affine transformation. The affine transformation dynamically generates channel-level scale vectors and offset vectors by embedding metadata and then linearly projecting them, and injects them into the feature stream with zero-initialized residuals. The conditionally modulated features are then processed by a classifier and upsampled to output a pixel-level change probability map, which is then thresholded to obtain a binary change detection result.

[0030] Additionally, it should be noted that conditional modulation is built on top of the change detection decoder in the form of a plug-in module. In at least one feature refinement stage of the decoder, feature-level affine transformations are applied to visual features with metadata embedding as a condition. The channel-level scale parameters and offset parameters of the affine transformation are dynamically generated by the metadata embedding through linear projection. The conditional modulation module can be used in conjunction with various change detection networks that adopt encoding and decoding structures, and the amount of additional parameters introduced does not exceed 0.5% of the basic network parameters.

[0031] In some specific embodiments, the feature-wise affine transformation (i.e., feature-wise linear modulation, or FiLM for short) is performed as follows: for the ... Phase Decoding features FiLM modulation is performed according to the following formula: Channel-level scale vector: ; Channel-level offset vector: ; FiLM modulation results: ; Zero-initialization residual injection ( Initialize to 0): ; in As a learnable scalar, initialized to 0. This indicates channel-wise multiplication (broadcast mechanism), ensuring that the conditional modulation module does not change the original feature distribution during the initial training phase.

[0032] Conditional modulation is performed independently at multiple resolution stages of the decoder, and the linear projection parameters of each stage are independent of each other; the modulation at lower resolution stages focuses on compensating for scale inconsistencies caused by differences in flight altitude, while the modulation at higher resolution stages focuses on compensating for feature shifts caused by optical distortion.

[0033] In addition, in this embodiment, the image feature extraction encoder is preferably a twin encoder based on the Visual State Space Model, which uses a multi-directional cross-scanning strategy for global sequence modeling, with a computational complexity of linear O(L); the decoder preferably adopts a progressive multi-stage structure, with each stage including parallel processing of global sequence modeling branches and local shape adaptive convolution branches; the limitation of the backbone network in this embodiment is only a preferred implementation method, and the core innovation of this invention (feature-level fusion mechanism of EXIF ​​physical parameters) is unrelated to the backbone network architecture. The method of this invention is also applicable to change detection networks with Transformer or convolutional neural networks as the backbone.

[0034] Finally, the model training in this embodiment uses a distortion-aware weighted loss function, which is a weighted combination of weighted binary cross-entropy loss and Lovász-Softmax loss; the weight of each pixel in the weighted binary cross-entropy loss is defined as: ; in The normalized radial distance from this pixel to the principal point of the image is used as a proxy variable for the intensity of optical distortion. This is a hyperparameter.

[0035] The total loss function is: ; Lovász-Softmax loss is used to directly optimize the intersection-over-union (IoU) metric.

[0036] Furthermore, when some parameters are missing in the EXIF ​​metadata, the missing components are filled with the mean of the corresponding parameters in the training set, and the system can still operate normally under the condition of incomplete metadata; when all EXIF ​​metadata is missing, the metadata embedding is replaced with an all-zero vector, and combined with the zero-initialization residual injection strategy, it automatically degenerates into a basic change detection mode without conditional modulation, and the inference results are basically consistent with the basic network.

[0037] In summary, this invention addresses the technical gap in existing change detection methods that completely ignore EXIF ​​metadata in UAV images. It systematically integrates the EXIF ​​physical parameters (flight altitude, attitude angle, focal length, etc.) from UAV image files and their associated flight records into a deep learning change detection framework, proactively compensating for two systemic problems: optical distortion from wide-angle lenses and scale inconsistencies caused by multi-temporal flight altitude differences. The method constructs a normalized metadata feature vector containing differential feature components, which is mapped to a high-dimensional embedding via a lightweight encoder. In at least one feature refinement stage of the decoder, dynamic channel-level modulation is applied conditioned on the metadata embedding through feature-level affine transformation, and zero-initialization residual injection ensures training stability. During training, a distortion-aware weighted loss with normalized radial distance as a proxy variable is used. The conditional modulation module is built in a pluggable manner, with additional parameters not exceeding 0.5% of the base network, allowing integration with various encoding / decoding change detection networks and supporting automatic fault-tolerant degradation when EXIF ​​is missing. On a self-built UAV panoramic change detection dataset, this method improves the F1 score by approximately 2.8 percentage points compared to the baseline, significantly improving robustness in scenarios with varying altitudes.

[0038] Therefore, the embodiments of the present invention have the following advantages over the prior art: 1) The EXIF ​​physical metadata system is systematically integrated into the deep learning change detection framework to make up for the shortcomings of existing methods in the utilization of physical parameters and provide a more targeted active compensation scheme.

[0039] 2) The differential feature vector explicitly encodes the changes in flight conditions in both time phases, which has strong physical interpretability and lower risk of overfitting than the attention-dependent implicit learning scheme in small sample remote sensing scenarios.

[0040] 3) The feature-level affine transformation applies dynamic compensation independently to each channel, which improves accuracy compared to the feature splicing method; the number of additional parameters does not exceed 0.5% of the basic network, and the engineering deployment cost is extremely low.

[0041] 4) Zero-initialization residual injection ensures that the behavior of the network is basically consistent with that of the base network in the early stage of training, and can be adapted to various encoding and decoding change detection networks (CNN, Transformer, SSM backbone are all applicable).

[0042] 5) The distortion-aware weighted loss uses radial distance as a proxy variable, which can achieve targeted enhanced supervision of edge distortion areas without camera calibration parameters, and is particularly effective for panoramic image scenes.

[0043] 6) The EXIF ​​missing fault tolerance mechanism ensures that the system can still operate normally when the metadata is incomplete, meeting the robustness requirements of industrial applications.

[0044] Finally, the embodiments of the present invention will be described in conjunction with actual engineering cases: Dual-time panoramic images of urban building areas taken using a DJI Matrice 300 RTK equipped with a Zenmuse P1 wide-angle camera (resolution 4000×3000 pixels, flight altitude 80~120m, ±15% altitude fluctuation, complete EXIF).

[0045] The EXIF ​​data is parsed using the exifread library to extract flight altitude H (GPSAltitude), GPS coordinates, focal length f (FocalLength), and image resolution; attitude angles are extracted by matching SRT flight log timestamps, and when unavailable, they are filled with the mean of the training set.

[0046] Construct a 9-dimensional normalized vector m, with the following 9 components: , , , , GPS offset , , Difference vector Also 9 dimensions (including the difference of all components), , and The three are concatenated to form a 27-dimensional extended vector (9+9+9=27).

[0047] MLP encoding: Linear(27→64)→BN→ReLU→Linear(64→128)→BN→ReLU, outputting a 128-dimensional embedding. The parameter count is approximately 0.02M. The VMamba twin encoder extracts four-scale features, which are then used to generate differential features via MSTFM. Conditional modulation layers are independently inserted into each of the three resolution stages of the decoder, injected with zero-initialized residuals. Training employs distortion-aware weighted loss (weighted loss). , AdamW optimizer, initial learning rate 1e-4, cosine annealing, 50 epochs, batch size 8.

[0048] Furthermore, universal adaptability verification was conducted by inserting the conditional modulation module of the present invention into ChangeFormer (Transformer backbone) and SNUNet (CNN backbone), respectively. Simulated scale difference experiments (1.1x scaling of T2 images) were performed on the LEVIR-CD public dataset. Both backbone networks achieved significant performance improvements, verifying the universality of the method of the present invention and its independence from the backbone.

[0049] Further verification of EXIF ​​missing fault tolerance was conducted under three conditions: (1) complete EXIF; (2) missing attitude angles (mean filling); (3) all missing EXIF ​​(zero vector replacement). The results showed reasonable stepwise degradation. Under condition (3), the performance of the method of the present invention was basically consistent with the basic change detection method, which verified the automatic degradation mechanism of zero initialization residual injection.

[0050] Reference Figure 2 A remote sensing image change detection system based on physical imaging parameter fusion includes: The first module 201 is used to acquire the first phase UAV remote sensing image and the second phase UAV remote sensing image, perform metadata parsing and encoding processing, and obtain metadata embedding features; The second module 202 is used to perform feature extraction processing on the first phase UAV remote sensing image and the second phase UAV remote sensing image based on the image feature extraction encoder to obtain multi-scale dual-temporal visual features. The third module 203 is used to take the metadata embedding features as conditional input, apply conditional modulation to the multi-scale dual-temporal visual features, and perform decoding output processing to obtain the UAV remote sensing change map.

[0051] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0052] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A remote sensing image change detection method based on physical imaging parameter fusion, characterized in that, Includes the following steps: Acquire UAV remote sensing images from the first and second time phases, perform metadata parsing and encoding processing, and obtain metadata embedding features; Based on the image feature extraction encoder, feature extraction processing is performed on the first phase UAV remote sensing image and the second phase UAV remote sensing image to obtain multi-scale dual-temporal visual features. Metadata embedding features are used as conditional inputs to apply conditional modulation to multi-scale dual-temporal visual features, and then the decoded output is processed to obtain a UAV remote sensing change map.

2. The remote sensing image change detection method based on fusion of physical imaging parameters according to claim 1, characterized in that, The step of acquiring the first-time-phase UAV remote sensing image and the second-time-phase UAV remote sensing image, performing metadata parsing and encoding processing to obtain metadata embedding features specifically includes: The first and second time-phase UAV remote sensing images are acquired, and the metadata of the physical imaging parameters is parsed to obtain the metadata of the remote sensing images. The metadata includes flight altitude, position information, lens focal length, attitude angle, sensor resolution and shooting time information. The metadata of the remote sensing image is normalized to obtain the first single-temporal metadata feature vector and the second single-temporal metadata feature vector; Obtain the dual-phase difference feature vector of the first single-phase metadata feature vector and the second single-phase metadata feature vector, and concatenate them with the first single-phase metadata feature vector and the second single-phase metadata feature vector to construct an extended metadata vector; The extended metadata vector is input into the metadata encoder for encoding to obtain the metadata embedding features.

3. The remote sensing image change detection method based on fusion of physical imaging parameters according to claim 2, characterized in that, The specific expression for calculating the dual-phase difference feature vector is as follows: ； In the above formulae, denotes a dual time phase difference feature vector, denotes a first single time phase metadata feature vector, denotes a second single time phase metadata feature vector.

4. The remote sensing image change detection method based on fusion of physical imaging parameters according to claim 3, characterized in that, Both the first and second single-phase metadata feature vectors include the log-normalized value of flight altitude, the trigonometric function encoded value of attitude angle, the encoded value of position offset, the normalized ground sampling distance value, and the encoded value of time period. The dual-phase difference feature vector includes the dual-phase flight altitude encoding difference, the ground sampling distance encoding difference, and the yaw angle encoding difference.

5. The remote sensing image change detection method based on physical imaging parameter fusion according to claim 4, characterized in that, The step of extracting features from the first and second temporal UAV remote sensing images based on the image feature extraction encoder to obtain multi-scale dual-temporal visual features specifically includes: The first-time-phase UAV remote sensing image and the second-time-phase UAV remote sensing image are input into the image feature extraction encoder, which includes an image feature extraction module, a shared weight VSSM encoder and an MSTFM fusion module. The image feature extraction module based on the image feature extraction encoder performs feature extraction processing on the first phase UAV remote sensing image and the second phase UAV remote sensing image to obtain the features of the first phase UAV remote sensing image and the features of the second phase UAV remote sensing image. Based on the image feature extraction encoder, the shared weight VSSM encoder performs weight-sharing encoding on the features of the first phase UAV remote sensing image and the features of the second phase UAV remote sensing image, to obtain the encoded features of the first phase UAV remote sensing image and the encoded features of the second phase UAV remote sensing image. The MSTFM fusion module based on the image feature extraction encoder performs feature fusion on the encoded first-temporal UAV remote sensing image features and the encoded second-temporal UAV remote sensing image features to obtain multi-scale dual-temporal visual features.

6. The remote sensing image change detection method based on physical imaging parameter fusion according to claim 5, characterized in that, The step of using metadata embedding features as conditional input to apply conditional modulation to multi-scale dual-temporal visual features and performing decoding output processing to obtain the UAV remote sensing change map specifically includes: Metadata embedding features are used as conditional inputs, and conditional modulation is applied to multi-scale dual-temporal visual features through feature-level affine transformation to obtain conditionally modulated remote sensing image features. The features of the conditionally modulated remote sensing image are sequentially classified and upsampled to obtain a pixel-level change probability map. Thresholding is performed on the pixel-level change probability map to obtain the UAV remote sensing change map.

7. The remote sensing image change detection method based on fusion of physical imaging parameters according to claim 6, characterized in that, The feature-level affine transformation represents the dynamic generation of channel-level scale vectors and channel-level offset vectors from metadata-embedded features via linear projection, and is injected into the feature stream with zero-initialized residuals, wherein: The expression for the channel-level scale vector is as follows: ； The expression for the channel-level offset vector is as follows: ； The specific expression for the modulation result of the feature-level affine transformation is as follows: ； The specific expression for injecting the feature stream using the zero-initialization residual method is as follows: ； In the above formula, Represents the channel-level scale vector. Represents the channel-level offset vector. Represents metadata embedding features, Indicates the first Phase Decoding features, Indicates a learnable scalar. This indicates channel-by-channel multiplication. This represents the linear projection weight matrix that generates the channel-level scale vector. This represents the linear projection bias that generates the channel-level scale vector. This represents the linear projection weight matrix that generates the channel-level offset vector. This represents the linear projection bias used to generate the channel-level offset vector. Represents the channel-level scale vector. Represents the channel-level offset vector. This indicates the decoding characteristics after modulation.

8. The remote sensing image change detection method based on fusion of physical imaging parameters according to claim 7, characterized in that, It also includes constructing a distortion-aware weighted loss function by combining the weighted binary cross-entropy loss function and the Lovász-Softmax loss function during the model training phase.

9. The remote sensing image change detection method based on fusion of physical imaging parameters according to claim 8, characterized in that, It also includes building a metadata missing fault tolerance mechanism. When some parameters in the metadata are missing, the missing components are filled with the mean of the corresponding parameters in the training set. When all metadata is missing, the metadata embedding is replaced with an all-zero vector.

10. A remote sensing image change detection system based on fusion of physical imaging parameters, characterized in that, Includes the following modules: The first module is used to acquire the first phase UAV remote sensing image and the second phase UAV remote sensing image, perform metadata parsing and encoding processing, and obtain metadata embedding features; The second module is used to perform feature extraction processing on the first phase UAV remote sensing image and the second phase UAV remote sensing image based on the image feature extraction encoder to obtain multi-scale dual-temporal visual features. The third module is used to take the metadata embedding features as conditional input, apply conditional modulation to the multi-scale dual-temporal visual features, and perform decoding output processing to obtain the UAV remote sensing change map.

Images