Unmanned aerial vehicle-satellite cross-aspect geo-location method based on frequency decomposition enhancement

Abstract

The application is suitable for the technical field of unmanned aerial vehicle observation, and provides an unmanned aerial vehicle-satellite cross-view angle geographic positioning method based on frequency decomposition enhancement, which comprises the following steps: acquiring a plurality of training samples; the training samples comprise unmanned aerial vehicle images and satellite images of the same geographic location; training a cross-view angle retrieval network by using the plurality of training samples; the cross-view angle retrieval network comprises a first feature extraction model and a second feature extraction model; extracting deep feature representation of each satellite image in a satellite image library by using the trained first feature extraction model, and constructing a satellite image feature library based on the extracted deep feature representation; extracting deep feature representation of a to-be-positioned unmanned aerial vehicle image by using the second feature extraction model, and performing similarity matching on the deep feature representation of the to-be-positioned unmanned aerial vehicle image and the deep feature representation of each satellite image in the satellite image feature library to obtain geographic location information of the to-be-positioned unmanned aerial vehicle image. The application can improve the accuracy of cross-view angle geographic positioning.

Description

Technical Field

[0001] This application belongs to the field of UAV observation technology, and in particular relates to a UAV-satellite cross-view geolocation method based on frequency decomposition enhancement. Background Technology

[0002] In recent years, with the rapid development of remote sensing platforms, UAV platforms, and deep learning technologies, cross-view matching of UAV images with satellite images possessing precise geographic coordinates to determine the current location of the UAV or the location of a target area has become an important research direction in the field of cross-view geolocation. The basic idea of ​​this method is to input the query image acquired by the UAV into the positioning system, then retrieve the best-matching candidate image from the satellite image database, and finally achieve indirect positioning based on the geographic coordinates corresponding to the candidate satellite image.

[0003] Existing cross-view geolocation technologies can generally be divided into two categories of application tasks. The first category is UAV image retrieval of satellite imagery, which uses UAV images as queries to search for corresponding geographic areas in a satellite image database. The second category is satellite image retrieval of UAV images, which uses satellite images to perform reverse matching on UAV observation images to support UAV navigation, historical trajectory retrieval, and area relocation. Most existing methods employ dual-branch deep neural networks or shared-encoding networks based on visual Transformers to extract unified embedding features from images from different viewpoints, and then achieve cross-view alignment through similarity measurement, contrastive learning, or classification supervision.

[0004] Existing research approaches primarily focus on spatial domain feature modeling. Early methods typically relied on global features to characterize the overall scene structure or improved fine-grained discrimination capabilities through local region modeling. Subsequent methods further introduced multi-scale fusion, block-level matching, attention mechanisms, and Transformer long-range dependency modeling to alleviate geometric differences and semantic offsets in cross-view scenarios. While these methods have improved cross-view matching performance to some extent, they still have the following significant shortcomings.

[0005] First, existing methods generally attribute viewpoint differences primarily to differences in spatial structure, while paying insufficient attention to frequency domain information. In reality, UAV images and satellite images differ significantly in imaging altitude, imaging angle, sensor characteristics, weather conditions, and lighting conditions. These differences not only manifest in spatial arrangement but also lead to systematic shifts in the responses of different frequency components. Matching only in the spatial domain is susceptible to interference from viewpoint misalignment, texture noise, and appearance variations.

[0006] Second, existing methods typically lack explicit measures and constraints for the differences between UAV feature distributions and satellite feature distributions. Even if images from different viewpoints originate from the same geographical location, their statistical distributions in the high-dimensional embedding space may still shift due to differences in imaging conditions. Without introducing a distribution alignment mechanism, relying solely on local alignment or sample-level similarity optimization can easily lead to insufficient generalization ability of the model under complex viewpoint changes, flight altitude changes, or weather disturbances.

[0007] Based on the above shortcomings, it is evident that the cross-view matching accuracy between UAV images and satellite images is currently low, resulting in low accuracy of cross-view geolocation and making it difficult to meet the requirements. Summary of the Invention

[0008] This application provides a UAV-satellite cross-view geolocation method based on frequency decomposition enhancement, which can solve the problem of low accuracy in cross-view geolocation.

[0009] This application provides a UAV-satellite cross-view geolocation method based on frequency decomposition enhancement, including: Acquire multiple training samples; each training sample includes drone images and satellite images of the same geographical location; The constructed cross-view retrieval network is trained using multiple training samples. The cross-view retrieval network includes a first feature extraction model and a second feature extraction model. Both the first and second feature extraction models include a VIT model, a frequency domain decomposition module, a cross-frequency interaction module, and a feature fusion module. The VIT model encodes the received data and outputs patch tokens and class tokens. The frequency domain decomposition module generates frequency domain features based on the patch tokens output by the VIT model and divides these features into low-frequency, mid-frequency, and high-frequency features. The cross-frequency interaction module processes the low-frequency, mid-frequency, and high-frequency features based on a frequency cross-attention mechanism to obtain low-frequency enhanced features and high-frequency enhanced features. The feature fusion module fuses the class token, low-frequency enhanced features, and high-frequency enhanced features to obtain a deep feature representation. During training, the total loss function consists of a maximum mean difference loss function, a contrastive learning loss function, and a supervised loss function. Satellite images of the training samples are input into the VIT model of the first feature extraction model, while drone images of the same training samples are input into the VIT model of the second feature extraction model. The first feature extraction model after training is used to extract deep feature representations of each satellite image in the satellite image database, and a satellite image feature database is constructed based on the extracted deep feature representations; The deep feature representation of the UAV image to be located is extracted using the trained second feature extraction model. The deep feature representation of the UAV image to be located is then matched with the deep feature representation of each satellite image in the satellite image feature library to obtain the geographical location information of the UAV image to be located.

[0010] Optionally, the frequency domain decomposition module is specifically used to perform the following steps: The patch tokens output by the VIT model are reconstructed into a two-dimensional spatial feature map according to the original patch space arrangement. Perform a two-dimensional fast Fourier transform on each channel of the two-dimensional spatial feature map to obtain the frequency domain features; Frequency domain features are divided into low-frequency features, mid-frequency features, and high-frequency features.

[0011] Optionally, the cross-frequency interaction module is specifically used to perform the following steps: The low-frequency enhancement characteristics are calculated using the following formula. and high frequency enhancement features : ; ; ; ; in, Represents the query matrix. This represents the query weight matrix. Indicates mid-frequency characteristics, This represents the key matrix corresponding to low-frequency features. Represents the key weight matrix. Indicates low-frequency characteristics. Represents a value matrix, Represents the value weight matrix, This represents the softmax function. Representing feature dimension, This represents the key matrix corresponding to high-frequency features. Indicates high-frequency characteristics.

[0012] Optionally, the feature fusion module is specifically used to perform the following steps: Average pooling is performed on low-frequency, mid-frequency, and high-frequency enhancement features respectively, and the average pooling results of low-frequency, mid-frequency, and high-frequency enhancement features are then concatenated. The splicing result is input into a fully connected network for processing to obtain frequency features; The class token is input into a fully connected network for processing to obtain global features; The frequency features and global features are concatenated, and the concatenated result is input into a fully connected network for processing to obtain a deep feature representation.

[0013] Optionally, the maximum mean difference loss function is: ; in, This represents the value of the maximum mean difference loss function. Indicates the number of training samples. Indicates the first Deep feature representation of drone images in training samples Represents the kernel mapping function. Indicates the first Deep feature representation of satellite images in training samples, Represents the regenerated nucleus Hilbert space.

[0014] Optionally, the contrastive learning loss function is: ; in, This represents the value of the contrastive learning loss function. Represents the cosine similarity function. Indicates the temperature coefficient. Indicates the first Deep feature representation of satellite images in training samples.

[0015] Optionally, the formula for calculating the total loss function is: ; in, This represents the value of the total loss function. This represents the value of the supervised loss function.

[0016] Optionally, the satellite image feature library stores the deep feature representation of each satellite image in the satellite image library, as well as the geographical location information of the satellite image corresponding to each deep feature representation.

[0017] Optionally, the deep feature representation of the UAV image to be located is compared with the deep feature representations of various satellite images in the satellite image feature library to obtain the geographic location information of the UAV image to be located, including: Calculate the cosine similarity between the deep feature representation of each satellite image in the satellite image feature library and the deep feature representation of the UAV image to be located; The geographic location information of the satellite image corresponding to the largest cosine similarity is used as the geographic location information of the UAV image to be located.

[0018] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, during the training process, a feature extraction model based on the VIT model and with frequency decomposition and cross-frequency interaction as its core extracts deep feature representations of UAV images and satellite images. The maximum mean difference loss function, contrastive learning loss function, and supervised loss function are introduced to construct feature difference measurement and joint supervised training, thereby simultaneously improving feature consistency, discriminability, and robustness. This enables the trained feature extraction model to overcome the differences in feature distribution across different perspectives. When using it to extract deep feature representations of the UAV image to be located, it can narrow the distance between the deep feature representations of the UAV image and the deep feature representations of the satellite image, accurately matching the satellite image corresponding to the UAV image to be located, and determining the geographical location of the UAV image to be located, thus improving the accuracy of cross-view geolocation. Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a UAV-satellite cross-view geolocation method based on frequency decomposition enhancement provided in an embodiment of this application. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0022] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0023] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0025] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] To address the low accuracy of current cross-view geolocation methods, this application provides a UAV-satellite cross-view geolocation method based on frequency decomposition enhancement. This method extracts deep feature representations from UAV and satellite images during training using a feature extraction model based on the VIT model and centered on frequency decomposition and cross-frequency interaction. It also introduces maximum mean difference loss, contrastive learning loss, and supervised loss to construct feature difference metrics and jointly supervise training, thereby simultaneously improving feature consistency, discriminability, and robustness. This allows the trained feature extraction model to overcome feature distribution differences across different viewpoints. When extracting deep feature representations from the UAV image to be located, it can narrow the distance between the deep feature representations of the UAV image and the satellite image, accurately matching the satellite image corresponding to the UAV image to be located, thus determining the geographical location of the UAV image and improving the accuracy of cross-view geolocation.

[0028] The following is an exemplary description of the UAV-satellite cross-view geolocation method based on frequency decomposition enhancement provided in this application, with reference to specific embodiments.

[0029] like Figure 1 As shown in the embodiments of this application, the UAV-satellite cross-view geolocation method based on frequency decomposition enhancement includes the following steps: Step 11: Obtain multiple training samples; each training sample includes drone images and satellite images of the same geographical location.

[0030] The aforementioned drone images were acquired by drones, while the satellite images were acquired by satellites. During the training phase, positive sample correspondences were established between the drone images and the satellite images based on their geographic location labels.

[0031] Understandably, to facilitate subsequent training, the aforementioned UAV and satellite images need to be preprocessed (i.e., the preprocessed images are used in the subsequent training process). This preprocessing includes image resizing, normalization, and enhancement.

[0032] Image size uniformity: The input image is adjusted to a uniform resolution to adapt to the image patch division method of the Visual Transformer (VIT model). Preferably, the input resolution is set to 392×392.

[0033] Normalization and Enhancement: Normalize UAV and satellite images; enhancement operations such as random cropping, flipping, or color perturbation can also be added to improve the model's generalization ability.

[0034] Step 12: Train the constructed cross-view retrieval network using multiple training samples.

[0035] The aforementioned cross-view retrieval network includes a first feature extraction model and a second feature extraction model with identical structures. Both the first and second feature extraction models include a VIT model, a frequency domain decomposition module, a cross-frequency interaction module, and a feature fusion module.

[0036] The VIT model is used to encode the received data and output patch tokens and class tokens; the frequency domain decomposition module is used to generate frequency domain features based on the patch tokens output by the VIT model, and divide the frequency domain features into low-frequency features, mid-frequency features, and high-frequency features; the cross-frequency interaction module is used to process the low-frequency features, mid-frequency features, and high-frequency features based on the frequency cross-attention mechanism to obtain low-frequency enhanced features and high-frequency enhanced features; the feature fusion module is used to fuse the class token, low-frequency enhanced features, and high-frequency enhanced features to obtain deep feature representations.

[0037] During training: satellite images of the training samples are input into the VIT model of the first feature extraction model, while drone images of the training samples are input into the VIT model of the second feature extraction model.

[0038] Specifically, the aforementioned VIT model is a Visual Transformer (also known as the DINOv2 Visual Transformer encoder), and the VIT models of the first and second feature extraction models share parameters. The VIT model performs forward encoding on UAV images and satellite images respectively. The VIT model includes a patch embedding layer, a class token, a position encoding layer, and multiple Transformer encoding blocks stacked along the depth direction. Each Transformer encoding block includes layer normalization, a multi-head self-attention sub-layer, residual connections, and a feedforward network sub-layer. For the input image, it is first divided into multiple non-overlapping image patches, and each image patch is mapped to a patch token; then, the class token is concatenated with each patch token, and position encoding is superimposed to form an input token sequence; then, the input token sequence is sequentially input into each Transformer encoding block for feature interaction and update; finally, the first output token is used as the class token, and the remaining tokens are used as patch tokens, which are used to represent the global semantic information and local region feature information of the image.

[0039] In some embodiments of this application, the frequency domain decomposition module described above is specifically used to perform steps 12.1 to 12.3: Step 12.1: Reconstruct the patch tokens output by the VIT model into a two-dimensional spatial feature map according to the original patch space arrangement. Specifically, scale reshaping can be used to process the patch tokens and reconstruct them into a two-dimensional spatial feature map.

[0040] Step 12.2: Perform a two-dimensional fast Fourier transform on each channel of the two-dimensional spatial feature map to obtain the frequency domain features. That is, the two-dimensional spatial feature map is mapped to the frequency domain through the two-dimensional fast Fourier transform to obtain the frequency domain features.

[0041] Step 12.3: Divide the frequency domain features into low-frequency features, mid-frequency features, and high-frequency features.

[0042] Specifically, frequency domain features can be divided into low-frequency features based on the frequency domain radius or a preset mask. Mid-frequency characteristics and high frequency characteristics : ; ; ; in, Represents frequency domain characteristics, This represents the preset low-frequency mask. This represents the preset intermediate frequency mask. This represents a preset high-frequency mask. This indicates element-wise multiplication.

[0043] In some embodiments of this application, the cross-frequency interaction module is specifically used to perform the following steps: The low-frequency enhancement features are calculated using the following formula. and high frequency enhancement features : ; ; ; ; in, Represents the query matrix. This represents the query weight matrix. Indicates mid-frequency characteristics, This represents the key matrix corresponding to low-frequency features. Represents the key weight matrix. Indicates low-frequency characteristics. Represents a value matrix, Represents the value weight matrix, This represents the softmax function. Representing feature dimension, This represents the key matrix corresponding to high-frequency features. Indicates high-frequency characteristics.

[0044] That is, the cross-frequency interaction module constructs a frequency cross-attention mechanism to realize the interaction between mid-frequency features and low-frequency features, and between mid-frequency features and high-frequency features, respectively, to obtain enhanced low-frequency features (i.e., low-frequency enhanced features) and enhanced high-frequency features (i.e., high-frequency enhanced features).

[0045] In some embodiments of this application, the feature fusion module described above is specifically used to perform steps 12.4 to 12.7: Step 12.4: Perform average pooling operations on the low-frequency enhancement features, mid-frequency features, and high-frequency enhancement features respectively, and then concatenate the average pooling results of the low-frequency enhancement features, mid-frequency features, and high-frequency enhancement features.

[0046] Step 12.5: Input the splicing result into a fully connected network for processing to obtain frequency features.

[0047] Specifically, frequency characteristics for: ; in, This represents a multilayer perceptron. This indicates a splicing operation. This indicates the average pooling operation.

[0048] Step 12.6: Input the class token into the fully connected network for processing to obtain global features.

[0049] Specifically, a multilayer perceptron can be used to process the class tokens output by the VIT model to obtain global features. .

[0050] Step 12.7: Concatenate the frequency features and global features, and input the concatenation result into the fully connected network for processing to obtain the deep feature representation.

[0051] Specifically, deep feature representation for: .

[0052] During training: The total loss function consists of the maximum mean difference loss function, the contrastive learning loss function, and the supervised loss function. Maximum mean difference loss, InfoNCE contrastive learning loss, and classification supervised loss are introduced into the deep feature representation to construct a joint training objective, complete model optimization, and thus achieve feature difference measurement and joint supervised optimization, overcoming the differences in feature distribution from different perspectives.

[0053] In some embodiments of this application, the maximum mean difference loss function is: ; in, This represents the value of the maximum mean difference loss function. Indicates the number of training samples. Indicates the first Deep feature representation of drone images in training samples Represents the kernel mapping function. Indicates the first Deep feature representation of satellite images in training samples, Represents the regenerated nucleus Hilbert space.

[0054] It is worth mentioning that the aforementioned maximum mean difference loss function is a feature difference metric constraint, which can explicitly reduce the statistical distribution difference between the deep features of UAVs and the deep features of satellites.

[0055] The contrastive learning loss function is as follows: ; in, This represents the value of the contrastive learning loss function. Represents the cosine similarity function. Indicates the temperature coefficient. Indicates the first Deep feature representation of satellite images in training samples.

[0056] It is worth mentioning that the contrastive learning loss function can ensure that drone and satellite features from the same geographical location are closely clustered in the embedding space, while samples from different locations are effectively separated (that is, ensuring that drone and satellite features from the same geographical location are close to each other in the embedding space, and keeping samples from different geographical locations far apart), thereby improving cross-view discrimination.

[0057] The aforementioned supervised loss function can be calculated using a conventional supervised loss function formula. It should be noted that, in some embodiments of this application, for any training sample, after the cross-view retrieval network outputs the deep feature representations of the UAV image and the satellite image, it processes these representations using classifiers to output category labels. Then, the supervised loss is calculated using the category labels output by the cross-view retrieval network and the ground truth labels of the UAV and satellite images in the training samples. Here, the ground truth labels are set when the training samples are acquired; for example, if the UAV and satellite images represent a certain type of building, the ground truth label is the building's identification number.

[0058] The formula for calculating the total loss function is as follows: ; in, This represents the value of the total loss function. This represents the value of the supervised loss function.

[0059] In some embodiments of this application, the cross-view retrieval network can be trained using conventional deep learning model training methods to optimize model parameters. It should be noted that the loss function used in the training process in this application is the aforementioned total loss function, which introduces maximum mean difference loss, InfoNCE contrastive learning loss, and classification supervision loss into the deep feature representation, enabling feature difference measurement and joint supervision optimization.

[0060] Step 13: Use the trained first feature extraction model to extract the deep feature representation of each satellite image in the satellite image library, and construct the satellite image feature library based on the extracted deep feature representation.

[0061] The aforementioned satellite image database stores a large number of satellite images to support subsequent positioning and can be a currently publicly available satellite image database. In some embodiments of this application, each satellite image in the satellite image database can be sequentially input into a trained first feature extraction model (i.e., the first feature extraction model in the trained cross-view retrieval network) for processing to obtain a deep feature representation of each satellite image in the satellite image database. Then, an association relationship is established between the deep feature representation and the geographical location information of the corresponding satellite image. Specifically, the satellite image feature database stores a deep feature representation of each satellite image in the satellite image database, as well as the geographical location information of the satellite image corresponding to each deep feature representation.

[0062] Step 14: Use the trained second feature extraction model to extract the deep feature representation of the UAV image to be located, and perform similarity matching between the deep feature representation of the UAV image to be located and the deep feature representation of each satellite image in the satellite image feature library to obtain the geographical location information of the UAV image to be located.

[0063] The above-mentioned drone images to be located are images collected by the drone, and the drone images to be located are drone images for which the geographical location needs to be determined.

[0064] In some embodiments of this application, the image of the UAV to be located can be input into a trained second feature extraction model (i.e., the second feature extraction model in the trained cross-view retrieval network) for processing to obtain a deep feature representation of the UAV image to be located.

[0065] In some embodiments of this application, the cosine similarity between the deep feature representation of each satellite image in the satellite image feature library and the deep feature representation of the UAV image to be located can be calculated respectively; then the geographical location information of the satellite image corresponding to the largest cosine similarity is used as the geographical location information of the UAV image to be located.

[0066] As an optional example, the cosine similarity between the deep feature representation of a satellite image in the satellite image feature library and the deep feature representation of the UAV image to be located can be used as the similarity between the two. After calculating all similarities, the geographical location information of the satellite image (which is a satellite image in the satellite image feature library) corresponding to the largest similarity value is used as the geographical location information of the UAV image to be located.

[0067] The following example illustrates the UAV-satellite cross-view geolocation method based on frequency decomposition enhancement provided in this application.

[0068] In this example, the public datasets University-1652 and SUES-200 are used as test data.

[0069] The University-1652 dataset is one of the most widely used datasets in the field of cross-view geolocation. This dataset contains 1652 building targets from 72 universities worldwide. Unlike traditional datasets that primarily focus on landmark buildings, University-1652 selects ordinary campus buildings as the localization targets, thus increasing the task difficulty and more closely resembling real-world application scenarios.

[0070] The SUES-200 dataset is a dataset specifically designed for UAV-satellite cross-view matching tasks, containing 24,120 pairs of UAV and satellite images. Collected in the area surrounding Shanghai University of Engineering Science, the dataset covers 200 different geographical locations and includes scene types such as campus buildings, parks, lakes, and public facilities, demonstrating rich environmental diversity.

[0071] Unlike University-1652, SUES-200 pays special attention to the perspective differences caused by altitude variations. At each sampling location, UAV images were acquired at four flight altitudes (150m, 200m, 250m, and 300m), with 50 UAV images corresponding to each altitude, thus forming a significant distribution of scale and perspective variations. Simultaneously, each location corresponds to one satellite image, forming cross-perspective matching pairs.

[0072] The performance comparison of cross-view geolocation using the method of this application with existing methods is shown in Tables 1, 2, and 3. Existing methods include Local Location Network (LPN), Feature Spatial Region Alignment (FSRA), Multi-Branch Fusion (MBF), Multi-Context Feature Gathering Map (MCCG), Geographic Sampling Model (Sample4Geo), and Dynamic Adaptive Calibration (DAC). In the tables, "ours" represents the method of this application. R@1 represents the first-response recall, and AP represents the average precision.

[0073] Table 1 Comparative Experiments with University-1652

[0074] Table 2 Comparison Experiment of SUES Dataset (UAV -> Satellite)

[0075] Table 3 Comparison Experiment of SUES Dataset (UAV -> Satellite)

[0076] The above comparison results show that the positioning accuracy of this application is significantly better than that of other existing methods.

[0077] In summary, the UAV-satellite cross-view geolocation method based on frequency decomposition enhancement provided in this application first extracts deep representations of UAV and satellite images using a visual Transformer, then decomposes the patch features in the frequency domain, dividing them into three sub-bands: low-frequency, mid-frequency, and high-frequency. Next, using mid-frequency structural information as a stable anchor point, a cross-frequency interaction mechanism is designed to allow low-frequency and high-frequency features to interact with mid-frequency features. After obtaining the enhanced deep features, a feature difference measurement mechanism is used to explicitly constrain the distribution differences between UAV and satellite features, and this is further optimized by combining contrastive learning and classification supervision, thereby simultaneously improving feature consistency, discriminativeness, and robustness. Specifically, this method has the following advantages: First, this application introduces a three-subband decomposition mechanism of low frequency, mid frequency, and high frequency, which enables differentiated processing of the semantic information carried by different frequency components. Compared with existing unified feature modeling methods, this application uses mid-frequency structural information as a cross-view stability anchor point and guides the interaction between low-frequency and high-frequency features and mid-frequency features through a cross-attention mechanism. This suppresses appearance perturbations while retaining detailed information effective for localization, thereby improving the stability of cross-view matching.

[0078] Second, this application introduces a feature difference measurement mechanism to explicitly constrain the feature distributions of UAVs and satellites. Compared to schemes that rely solely on sample-level similarity optimization, this application reduces cross-viewpoint offset at the statistical distribution level through maximum mean difference loss, thereby improving feature consistency and model generalization ability.

[0079] Third, this application employs a joint optimization approach combining feature difference measurement, contrastive learning, and classification supervision, while simultaneously ensuring the consistency, discriminativeness, and geographic location distinguishability of features across different perspectives. Compared to a single loss training strategy, this application can more stably learn deep embedding representations suitable for retrieval and localization tasks.

[0080] Fourth, this application achieves superior performance results compared to existing mainstream methods on publicly available datasets, validating the effectiveness of the technical solution. On the University-1652 dataset, this application achieves R@1=95.59% and AP=96.11% in UAV satellite retrieval tasks, outperforming the current state-of-the-art (SOTA) method DAC (R@1=94.67% and AP=95.50%); and in satellite retrieval UAV tasks, it achieves R@1=97.20% and AP=94.67%, again outperforming the DAC method (R@1=96.43% and AP=93.79%).

[0081] Fifth, this application maintains excellent performance under different flight altitude conditions, demonstrating good cross-scale generalization ability. On the SUES-200 dataset, in UAV satellite retrieval missions, this application achieves R@1 / AP results of 97.00 / 97.68, 97.63 / 98.13, 98.50 / 99.11, and 99.02 / 99.27 at four altitudes: 150m, 200m, 250m, and 300m, respectively; in satellite retrieval UAV missions, it achieves R@1 / AP results of 99.25 / 96.39, 98.75 / 97.72, 99.65 / 98.68, and 100 / 99.19, respectively, indicating that this application has strong adaptability to changes in UAV altitude.

[0082] Sixth, ablation experiments have verified that each component module makes a clear contribution to performance improvement. Compared with the basic model, the complete model improves the R@1 and AP metrics in both retrieval directions, indicating that the frequency-aware interaction mechanism and the feature difference measurement mechanism can work together to improve cross-view localization accuracy and robustness.

[0083] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A UAV-satellite cross-view geolocation method based on frequency decomposition enhancement, characterized in that, include: Obtain multiple training samples; Each training sample includes drone images and satellite images of the same geographical location; The constructed cross-view retrieval network is trained using the multiple training samples. The cross-view retrieval network includes a first feature extraction model and a second feature extraction model. Both the first and second feature extraction models include a VIT model, a frequency domain decomposition module, a cross-frequency interaction module, and a feature fusion module. The VIT model is used to encode the received data and output patch tokens and class tokens. The frequency domain decomposition module is used to generate frequency domain features based on the patch tokens output by the VIT model and divide the frequency domain features into low-frequency features, mid-frequency features, and high-frequency features. The cross-frequency interaction module is used to process the low-frequency features, mid-frequency features, and high-frequency features based on a frequency cross-attention mechanism to obtain low-frequency enhanced features and high-frequency enhanced features. The feature fusion module is used to fuse the class token, low-frequency enhanced features, and high-frequency enhanced features to obtain a deep feature representation. During training, the total loss function consists of the maximum mean difference loss function, the contrastive learning loss function, and the supervised loss function. The satellite images of the training samples are input into the VIT model of the first feature extraction model, while the drone images of the training samples are input into the VIT model of the second feature extraction model. The first feature extraction model after training is used to extract deep feature representations of each satellite image in the satellite image database, and a satellite image feature database is constructed based on the extracted deep feature representations; The deep feature representation of the UAV image to be located is extracted using the trained second feature extraction model. The deep feature representation of the UAV image to be located is then matched with the deep feature representation of each satellite image in the satellite image feature library to obtain the geographical location information of the UAV image to be located.

2. The UAV-satellite cross-view geolocation method according to claim 1, characterized in that, The frequency domain decomposition module is specifically used to perform the following steps: The patch tokens output by the VIT model are reconstructed into a two-dimensional spatial feature map according to the original patch space arrangement. Perform a two-dimensional fast Fourier transform on each channel of the two-dimensional spatial feature map to obtain frequency domain features; The frequency domain features are divided into low-frequency features, mid-frequency features, and high-frequency features.

3. The UAV-satellite cross-view geolocation method according to claim 1, characterized in that, The cross-frequency interaction module is specifically used to perform the following steps: The low-frequency enhancement characteristics are calculated using the following formula. and high frequency enhancement features : ； ； ； ； in, Represents the query matrix. This represents the query weight matrix. Indicates mid-frequency characteristics, This represents the key matrix corresponding to low-frequency features. Represents the key weight matrix. Indicates low-frequency characteristics. Represents a value matrix, Represents the value weight matrix, This represents the softmax function. Representing feature dimension, This represents the key matrix corresponding to high-frequency features. Indicates high-frequency characteristics.

4. The UAV-satellite cross-view geolocation method according to claim 1, characterized in that, The feature fusion module is specifically used to perform the following steps: Average pooling is performed on the low-frequency enhancement features, the mid-frequency features, and the high-frequency enhancement features respectively, and the average pooling results of the low-frequency enhancement features, the mid-frequency features, and the high-frequency enhancement features are then concatenated. The splicing result is input into a fully connected network for processing to obtain frequency features; The class token is input into a fully connected network for processing to obtain global features; The frequency features and the global features are concatenated, and the concatenation result is input into a fully connected network for processing to obtain a deep feature representation.

5. The UAV-satellite cross-view geolocation method according to claim 1, characterized in that, The maximum mean difference loss function is: ； in, This represents the value of the maximum mean difference loss function. Indicates the number of training samples. Indicates the first Deep feature representation of drone images in training samples Represents the kernel mapping function. Indicates the first Deep feature representation of satellite images in training samples, Represents the regenerated nucleus Hilbert space.

6. The UAV-satellite cross-view geolocation method according to claim 5, characterized in that, The contrastive learning loss function is: ； in, This represents the value of the contrastive learning loss function. Represents the cosine similarity function. Indicates the temperature coefficient. Indicates the first Deep feature representation of satellite images in training samples.

7. The UAV-satellite cross-view geolocation method according to claim 6, characterized in that, The formula for calculating the total loss function is: ； in, This represents the value of the total loss function. This represents the value of the supervised loss function.

8. The UAV-satellite cross-view geolocation method according to claim 1, characterized in that, The satellite image feature library stores the deep feature representation of each satellite image in the library, as well as the geographical location information of the satellite image corresponding to each deep feature representation.

9. The UAV-satellite cross-view geolocation method according to claim 8, characterized in that, The deep feature representation of the image of the UAV to be located is matched with the deep feature representation of each satellite image in the satellite image feature library to obtain the geographical location information of the image of the UAV to be located, including: Calculate the cosine similarity between the deep feature representation of each satellite image in the satellite image feature library and the deep feature representation of the UAV image to be located; The geographic location information of the satellite image corresponding to the largest cosine similarity is used as the geographic location information of the UAV image to be located.

Images