Unmanned aerial vehicle cross-view geographic positioning method based on state space model

Abstract

The invention discloses an unmanned aerial vehicle cross-view-angle geographic positioning method based on a state space model, and relates to the technical field of geographic positioning methods, and the method is characterized in that the method comprises the following steps: S1, constructing a state space feature extraction module, and extracting features irrelevant to view angles from an unmanned aerial vehicle and a satellite image; s2, constructing a state space feature cross fusion module to obtain fused cross-view features; s3, building an unmanned aerial vehicle positioning network, and regressing the geographic coordinates of the unmanned aerial vehicle through a stacking feature extraction and feature cross fusion module; and S4, expanding satellite and unmanned aerial vehicle data, and updating unmanned aerial vehicle positioning network parameters by minimizing joint loss. The technical problem to be solved by the invention is to provide the unmanned aerial vehicle cross-view geographic positioning method based on the state space model, global features in an unmanned aerial vehicle and a satellite image are extracted through the state space model, cross-view features are fused by using a local attention mechanism, and geographic coordinates of the unmanned aerial vehicle are regressed on the satellite image.

Description

Technical Field

[0001] The present invention relates to the technical field of geolocation methods, and in particular to a cross-viewpoint geolocation method for an unmanned aerial vehicle (UAV) based on a state space model. Background Art

[0002] With the rapid development of drone technology, it has found widespread application in fields such as agriculture, environmental monitoring, and modern warfare. To achieve precise flight navigation and locate ground targets, drones typically rely on satellite signals provided by the Global Positioning System (GPS) to obtain real-time location information. However, in complex environments such as urban canyons, mountainous areas, or densely built-up areas, GPS signals are easily blocked or completely lost, resulting in reduced drone positioning accuracy and even the risk of loss of control.

[0003] Therefore, cross-view geolocation technology has gradually become an effective alternative to GPS and has attracted widespread attention. Existing cross-view geolocation methods can be divided into two main categories: one is based on image retrieval, which cuts satellite images into small pieces and retrieves the parts that are most similar to drone images to determine the drone's geographic coordinates; the other is based on keypoint regression, which extracts and fuses features from satellite and drone images to generate a keypoint heatmap of the satellite image and uses the high-value points in the heatmap to determine the drone's geographic location. Although existing methods obtain geographic coordinates by matching drone and satellite images from different perspectives, they still have shortcomings in extracting global features and fail to fully utilize the image's contextual information, which directly affects positioning accuracy.

[0004] Existing cross-viewpoint geolocation methods for drones match drone images with satellite imagery from different viewpoints and use the satellite imagery's location information to determine the drone's geographic coordinates. To improve positioning accuracy, satellite imagery often requires fine segmentation. However, this approach typically only achieves block-level accuracy and cannot provide pixel-level precision, leading to errors in positioning results. Furthermore, satellite image segmentation can produce a large number of duplicate image regions, resulting in data redundancy and increased storage and computing resource requirements. Furthermore, existing methods lack global feature extraction and fail to fully utilize the contextual information of satellite and drone imagery, compromising positioning accuracy. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a cross-viewpoint geolocation method for drones based on a state-space model. This method uses the state-space model to extract view-independent features from drone and satellite imagery. A state-space feature cross-fusion module is then used to extract view-independent features and regress the drone's geographic coordinates from the satellite imagery. This method can provide accurate real-time positioning data even when GPS signals are weak or absent, enabling precise drone navigation.

[0006] The present invention adopts the following technical solutions to achieve the invention objectives: A cross-viewpoint geolocation method for an unmanned aerial vehicle (UAV) based on a state-space model, comprising the following steps: S1: Construct a state-space feature extraction module to extract view-independent features from UAV and satellite images respectively; S2: Construct a state-space feature cross-fusion module to obtain fused cross-view features; S3: Build a UAV positioning network and regress the UAV’s geographic coordinates; S31: Stacked state-space feature extraction and cross-fusion modules to extract deep semantic features from drone and satellite images; S32: Extract deep semantic features through the decoder and use the positioning head to determine the geographic coordinates of the drone; S4: Model optimization, expanding satellite and drone data, and updating the network parameters of S3 by minimizing the joint loss; S41: data expansion; S42: Model optimization.

[0007] As a further limitation of this technical solution, the specific steps of S1 are: S11: Input the data into the state space feature extraction module. Each module contains two Mamba blocks with the same structure, which are used to process UAV and satellite data respectively. S12: The input data of each Mamba block is first embedded through the normalization layer and then divided into two completely consistent information streams; S13: Information flow 1 passes through a linear layer and a weighted linear activation function in sequence; S14: Information flow 2 passes through another linear layer, a 3×3 depthwise separable convolutional layer, and a 2D selective scanning module. The process of the 2D selective scanning module includes three steps: S141: Directional expansion, for the input feature map of the two-dimensional selective scanning module , expanded into a one-dimensional sequence along four directions: (1); in: Represents a function that expands a two-dimensional feature map into a one-dimensional sequence; Indicates the A sequence that expands in two directions; and Indicates scanning from upper left to lower right and from lower right to upper left in the horizontal direction; and Indicates scanning from upper left to lower right and from lower right to upper left in the vertical direction; S142: State space modeling, each sequence For causal sequence modeling, the discrete state space model is expressed as: (2); in: Indicates the first The internal hidden state at each moment; represents the state transition matrix; represents the input transformation matrix; Indicates the first The input vector at each moment; represents the output features; represents the mapping matrix; and Discretization of state-space models from continuous form; (3); in: and represents the original parameters of the continuous state space model; represents the time scale parameter used to map continuous parameters; represents the matrix exponential function; represents the identity matrix; S143: Reconstruction fusion, output sequence in each direction Reconstructed into feature map structure: (4); in: Indicates restoring a one-dimensional sequence to a two-dimensional feature map by direction; Represents the restored two-dimensional feature map;

[0008] Finally, the four directional outputs are fused into the output feature map of the two-dimensional selective scanning module: (5); in, Represents element-wise addition fusion function;

[0009] S15: Fusing the outputs of data stream 1 and data stream 2 by element-by-element multiplication; S16: The fused features pass through a linear layer to generate the output features of the Mamba block: (6); in: and Represent the input and output of the Mamba block respectively; Represents a pixel-by-pixel multiplication operation; represents a linear layer; Represents the linear layer operation of data flow one through; Represents the linear layer operation of data flow two; represents the weighted linear activation function; represents the normalization layer; represents a two-dimensional selective scanning module; represents a depth-wise separable convolutional layer;

[0010] As a further limitation of this technical solution, the specific steps of S2 are: S21: Extract features from drone images using the state space feature extraction module of S1 , extracting features from satellite images , and are input into a linear layer and a depth-separable convolution layer with a convolution kernel of 3×3 respectively to extract the local semantic features of the drone and satellite local semantic features : (7); in, and Represents the linear layers for processing drone and satellite data respectively; and Denote the depthwise separable convolutional layers for processing UAV and satellite data, respectively; S22: and Input into the cross-mamba block so that the cross-view features contain rich interaction information; The cross-mamba block process consists of two steps: S221: Obtain the state transfer matrix and transformation matrix by discretizing the state space model: ; (8); in, and Transfer matrix and transformation matrix representing the discretization of the UAV; and Transfer matrix and transformation matrix representing satellite discretization; and Represents the original parameters of the UAV continuous state space model; and Represents the original parameters of the satellite continuous state space model; and represent the time scale parameters of UAV and satellite data, respectively; S222: Cross-perspective causal sequence modeling, and Expanded into one-dimensional sequences respectively, the output of the cross-mamba block is expressed as: (9); in: and Represents the first The input vector at each moment; and represent the UAV and satellite features output by the cross-Mamba block, respectively; Indicates that the drone sequence is in The hidden state vector at the moment; Indicates that the satellite sequence is The hidden state vector at the moment; Represents the satellite observation matrix, which is used to Cross-view mapping into output features ; Represents the observation matrix of the UAV, which is used to Cross-view mapping into output features ; S23: and The features of the drone and satellite are fused by sending them to the normalization layer and the linear layer respectively, and performing element-wise multiplication: (10); in: Represents the fused cross-view features; represents the linear layer used to process drone features; represents the linear layer used to process satellite features; represents the normalization layer used to process drone features; Represents a normalization layer used to process satellite features.

[0011] As a further limitation of this technical solution, the specific steps of S3 are: S31: For the input drone image and satellite imagery , first reduce their spatial resolution by half through bilinear interpolation method, then, the reduced resolution UAV images and satellite imagery Send to backbone network; S311: In the first stage of the backbone network, the features extracted by the twin feature extraction module are represented as: (11); in: and They represent the Mamba blocks used for feature extraction of UAV and satellite images in the first stage, respectively; and Represent the UAV and satellite features extracted in the first stage respectively; S312: The cross-view features extracted by the state space feature cross fusion module are expressed as: (12); in: Represents the operation of the first-stage state space feature cross-fusion module; S313: The view-independent features extracted in the first stage and , as the input of the next stage, used to extract deep semantic features, Cross-view features fused in multiple stages Expressed as: (13); in, and Respectively represent The Mamba block is used for feature extraction of UAV and satellite images; Indicates the Operation of the stage state space feature cross fusion module; and Respectively represent UAV and satellite features extracted in the first stage; S32: The cross-view features extracted in the four stages are input to the positioning neck. The obtained features are expressed as follows: (14); in: Indicates a 2x upsampling operation; Indicates the A 1×1 convolutional layer; Indicates channel-based splicing operation; S33: Position the neck features Input to the positioning head, through the linear layer regression heat map, its resolution is the same as the satellite image Same, the maximum value in the heat map corresponds to the satellite image The coordinates of the image are the drone image The center point coordinates of the drone are obtained to achieve geographic positioning.

[0012] As a further limitation of this technical solution, the specific steps of S4 are: S41: Satellite imagery and drone imagery Perform data augmentation; S411: Expanding with random pruning strategies ,make and Satellite images The length and width, first of all, in satellite images Find drone imagery in Center coordinates; then, take the coordinates as the center and The random cropping length and width are and Partial image of ; Finally, in the local image Find a random coordinate and construct a structure with the length and width as the center. Augmented satellite imagery ; S412: Expanding with foreground mask strategy , first construct a Then, use the mask to map the drone image Finally, the pixels in the covered area are set to 0 to obtain the expanded drone image. ; S42: Model optimization; S421: Standard heat map supervision, original drone image and expanded satellite imagery As the input of the network in step S3, the heat map of the drone image is predicted , using the prediction heat map and the real heat map Calculate weight balance loss , as shown below: (15); in: Represents the index of each pixel position in the image; Represents the weight factor for balancing positive and negative samples; Indicates the The actual thermal value corresponding to each pixel on the label; Indicates the corresponding The predicted thermal value of each pixel; S422: Contextual reasoning supervision of occluded areas, expanding satellite images and drone imagery As the network input in step S3, obtain the predicted heat map , using the prediction heat map and prediction heatmap Calculate mean squared error loss , as shown below: (16); in: and Represent the length and width of the heat map respectively; S423: Jointly optimize the objective function. Combine formulas (15) and (16) to calculate the joint loss function: (17); in: represents the hyperparameter used to balance the weights of the two losses.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are: 1. The purpose of this invention is to propose a robust UAV geolocation solution. This solution effectively extracts global features from satellite and UAV imagery by introducing a state-space model, and utilizes a cross-fusion module to extract cross-view features, thereby accurately regressing the UAV's geographic coordinates. Compared with existing methods, this invention eliminates the need for segmenting satellite imagery, thereby reducing data redundancy, and significantly improves the accuracy of UAV geographic coordinate regression by mining global features. This solution can provide UAVs with more reliable positioning and navigation capabilities, enabling precise positioning even in complex environments where GPS signals are weak or absent.

[0014] 2. This paper proposes a novel UAV positioning method that constructs a cross-view information interaction network based on a state-space model. This method introduces a state-space modeling mechanism for the first time in UAV geolocation tasks. It achieves efficient modeling of global image information with linear computational complexity, effectively alleviating the limited receptive field of traditional convolutional neural networks and the high computational complexity of transformer models. A state-space feature extraction module is constructed to extract structurally consistent feature representations from UAV and satellite imagery, enhancing feature matching between the two. A state-space cross-fusion module is designed to achieve fine-grained region alignment between cross-modal images, improving the accuracy of heatmap positioning. A pyramid structure is used to fuse semantic information at different levels to enhance the perception of multi-scale objects in complex scenes. A foreground mask-based data augmentation strategy is proposed to occlude the foreground areas of UAV images during training, guiding the model to focus on more contextual cues and improving positioning robustness under complex conditions such as occlusion and angle changes. This method demonstrates significant advantages in UAV cross-view positioning tasks, combining efficient modeling, precise matching, and strong generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the process of the present invention.

[0016] Figure 2 Schematic diagram of the Mamba feature extraction block structure of the present invention.

[0017] Figure 3 It is a schematic diagram of the flow of the two-dimensional selective scanning module of the present invention.

[0018] Figure 4 Schematic diagram of the cross-feature fusion module structure of the present invention.

[0019] Figure 5 This is a schematic diagram of the UAV positioning network structure of the present invention. DETAILED DESCRIPTION

[0020] A specific embodiment of the present invention is described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.

[0021] The present invention proposes a cross-view geolocation method for unmanned aerial vehicles. The method constructs a UAV positioning network based on a state-space model, which mainly includes two core stages: feature extraction and feature fusion. In the feature extraction stage, a state-space feature extraction module is designed, and the Mamba block is used to extract discriminative view-independent features from UAV images and satellite images. In the feature fusion stage, a state-space feature cross-fusion module is designed to effectively enhance the cross-view correlation modeling capability. In the positioning neck and positioning head, a satellite image heat map is generated by fusing multi-scale features from different stages, and the peak position of the heat map is used as the predicted geographic coordinates of the UAV image to achieve precise positioning. In addition, a foreground mask data enhancement strategy is proposed to guide the model to learn richer spatial clues from the context, thereby improving the robustness and positioning accuracy in complex scenes.

[0022] The present invention comprises the following steps: S1: Construct a state-space feature extraction module to extract view-independent features from UAV and satellite images respectively.

[0023] Because drone and satellite imagery differ significantly in perspective, resolution, and lighting, a key challenge in the feature extraction phase is minimizing the differences between similar information. This requires the feature extraction network to be able to maintain feature similarity. To this end, a state-space feature extraction module is constructed.

[0024] The state space feature extraction module includes two parallel Mamba blocks with similar structures and parameter sharing, which are used to mine view-independent features and ensure that they remain highly consistent when identifying similar regions; Figure 2 Showing the structure of the Mamba block.

[0025] The specific steps of S1 are: S11: Input the data into the state space feature extraction module. Each module contains two Mamba blocks with the same structure, which are used to process UAV and satellite data respectively. S12: The input data of each Mamba block is first embedded through the normalization layer and then divided into two completely consistent information streams; S13: Information flow 1 passes through a linear layer and a weighted linear activation function in sequence; S14: Information flow 2 passes through another linear layer, a 3×3 depthwise separable convolutional layer, and a 2D selective scanning module. like Figure 3 As shown, the two-dimensional selective scanning module includes three steps: 141: Directional expansion, for the input feature map of the two-dimensional selective scanning module , expanded into a one-dimensional sequence along four directions: (1); in: Represents a function that expands a two-dimensional feature map into a one-dimensional sequence; Indicates the A sequence that expands in two directions; and Indicates scanning from upper left to lower right and from lower right to upper left in the horizontal direction; and Indicates scanning from upper left to lower right and from lower right to upper left in the vertical direction; S142: State space modeling, each sequence For causal sequence modeling, the discrete state space model is expressed as: (2); in: Indicates the first The internal hidden state at each moment; represents the state transition matrix; represents the input transformation matrix; Indicates the first The input vector at each moment; represents the output features; represents the mapping matrix; and Discretization of state-space models from continuous form; (3); in: and represents the original parameters of the continuous state space model; represents the time scale parameter used to map continuous parameters; represents the matrix exponential function; represents the identity matrix; S143: Reconstruction fusion, output sequence in each direction Reconstructed into feature map structure: (4); in: Indicates restoring a one-dimensional sequence to a two-dimensional feature map by direction; Represents the restored two-dimensional feature map.

[0026] Finally, the four directional outputs are fused into the output feature map of the two-dimensional selective scanning module: (5); in, Represents element-wise addition of fused functions.

[0027] S15: Fusing the outputs of data stream 1 and data stream 2 by element-by-element multiplication; S16: The fused features pass through a linear layer to generate the output features of the Mamba block: (6); in: and Represent the input and output of the Mamba block respectively; Represents a pixel-by-pixel multiplication operation; represents a linear layer; Represents the linear layer operation of data flow one through; Represents the linear layer operation of data flow two; represents the weighted linear activation function; represents the normalization layer; represents a two-dimensional selective scanning module; represents a depth-wise separable convolutional layer; S2: Construct a state-space feature cross-fusion module to obtain fused cross-view features.

[0028] After the state space feature extraction module, a state space feature cross fusion module is connected; Figure 4 As shown in Figure 2, the state-space feature cross-fusion module aggregates features from different perspectives to obtain fused view-independent features.

[0029] The specific steps of S2 are: S21: Extract features from drone images using the state space feature extraction module of S1 , extracting features from satellite images , and are input into a linear layer and a depth-separable convolution layer with a convolution kernel of 3×3 respectively to extract the local semantic features of the drone and satellite local semantic features : (7); in, and Represents the linear layers for processing drone and satellite data respectively; and Denote the depthwise separable convolutional layers for processing UAV and satellite data, respectively; S22: and Input into the cross-mamba block so that the cross-view features contain rich interaction information; The cross-mamba block process consists of two steps: S221: Obtain the state transfer matrix and transformation matrix by discretizing the state space model: ; (8); in, and Transfer matrix and transformation matrix representing the discretization of the UAV; and Transfer matrix and transformation matrix representing satellite discretization; and Represents the original parameters of the UAV continuous state space model; and Represents the original parameters of the satellite continuous state space model; and represent the time scale parameters of UAV and satellite data, respectively; S222: Cross-perspective causal sequence modeling, and Expanded into one-dimensional sequences respectively, the output of the cross-mamba block is expressed as: (9); in: and Represents the first The input vector at each moment; and represent the UAV and satellite features output by the cross-Mamba block, respectively; Indicates that the drone sequence is in The hidden state vector at the moment; Indicates that the satellite sequence is The hidden state vector at the moment; Represents the satellite observation matrix, which is used to Cross-view mapping into output features ; Represents the observation matrix of the UAV, which is used to Cross-view mapping into output features ; S23: and The features of the drone and satellite are fused by sending them to the normalization layer and the linear layer respectively, and performing element-wise multiplication: (10); in: Represents the fused cross-view features; represents the linear layer used to process drone features; represents the linear layer used to process satellite features; represents the normalization layer used to process drone features; Represents a normalization layer used to process satellite features.

[0030] S3: Build a UAV positioning network and regress the UAV coordinates by stacking feature extraction and fusion modules.

[0031] like Figure 5 As shown in the figure, the constructed drone positioning network includes a backbone network, a positioning neck, and a positioning head; the backbone network extracts cross-view features from different scales through four stages; the positioning neck fuses features from the four stages of the backbone network; the positioning head regresses the heat map to determine the position of the drone center point on the satellite image.

[0032] The specific steps of S3 are: S31: For the input drone image and satellite imagery , first reduce their spatial resolution by half through bilinear interpolation method, then, the reduced resolution UAV images and satellite imagery Send to backbone network; S311: In the first stage of the backbone network, the features extracted by the twin feature extraction module are represented as: (11); in: and They represent the Mamba blocks used for feature extraction of UAV and satellite images in the first stage, respectively; and Represent the UAV and satellite features extracted in the first stage respectively; S312: The cross-view features extracted by the state space feature cross fusion module are expressed as: (12); in: Represents the operation of the first-stage state space feature cross-fusion module; S313: The view-independent features extracted in the first stage and , as the input of the next stage, used to extract deep semantic features, Cross-view features fused in multiple stages Expressed as: (13); in, and Respectively represent The Mamba block is used for feature extraction of UAV and satellite images; Indicates the Operation of the stage state space feature cross fusion module; and Respectively represent UAV and satellite features extracted in the first stage; S32: The cross-view features extracted in the four stages are input to the positioning neck. The obtained features are expressed as follows: (14); in: Indicates a 2x upsampling operation; Indicates the A 1×1 convolutional layer; Indicates channel-based splicing operation; S33: Position the neck features Input to the positioning head, through the linear layer regression heat map, its resolution is the same as the satellite image Same, the maximum value in the heat map corresponds to the satellite image The coordinates of the image are the drone image The center point coordinates of the drone are obtained to achieve geographic positioning.

[0033] S4: Model optimization, expanding satellite and drone data, and updating the network parameters of S3 by minimizing the joint loss.

[0034] The specific steps of S4 are: S41: Satellite imagery and drone imagery Perform data augmentation; S411: Expanding with random pruning strategies ,make and Satellite images The length and width, first of all, in satellite images Find drone imagery in Center coordinates; then, take the coordinates as the center and The random cropping length and width are and Partial image of ; Finally, in the local image Find a random coordinate and construct a structure with the length and width as the center. Augmented satellite imagery ; S412: Expanding with foreground mask strategy , first construct a Then, use the mask to map the drone image Finally, the pixels in the covered area are set to 0 to obtain the expanded drone image. ; S42: Model optimization; S421: Standard heat map supervision, original drone image and expanded satellite imagery As the input of the network in step S3, the heat map of the drone image is predicted , using the prediction heat map and the real heat map Calculate weight balance loss , as shown below: (15); in: Represents the index of each pixel position in the image; Represents the weight factor for balancing positive and negative samples; Indicates the The actual thermal value corresponding to each pixel on the label; Indicates the corresponding The predicted thermal value of each pixel; S422: Contextual reasoning supervision of occluded areas, expanding satellite images and drone imagery As the network input in step S3, obtain the predicted heat map , using the prediction heat map and prediction heatmap Calculate mean squared error loss , as shown below: (16); in: and Represent the length and width of the heat map respectively; S423: Jointly optimize the objective function. Combine formulas (15) and (16) to calculate the joint loss function: (17); in: represents the hyperparameter used to balance the weights of the two losses.

[0035] Parameter update. In each training iteration, minimize the joint loss The backpropagation algorithm is used to update the parameters of the drone positioning network based on the state-space model. An early stopping strategy is used to determine whether the model has converged, and training stops when convergence is achieved. In the application scenario, the drone positioning network with trained parameters will be used for drone geolocation tasks.

[0036] The above disclosure is only a specific embodiment of the present invention, but the present invention is not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. A cross-view geolocation method for UAVs based on a state-space model, characterized in that: The following steps are involved: S1: Construct a state-space feature extraction module to extract view-independent features from UAV and satellite images respectively; S2: Construct a state-space feature cross-fusion module to obtain fused cross-view features; S3: Build a UAV positioning network and regress the UAV’s geographic coordinates; S4: Model optimization, expanding satellite and drone data, and updating the network parameters of S3 by minimizing the joint loss; S41: data expansion; S42: Model optimization.

2. The cross-viewpoint geolocation method for UAVs based on a state-space model according to claim 1, characterized in that: The specific steps of S1 are: S11: Input the data into the state space feature extraction module. Each module contains two Mamba blocks with the same structure, which are used to process UAV and satellite data respectively. S12: The input data of each Mamba block is first embedded through the normalization layer and then divided into two completely consistent information streams; S13: Information flow 1 passes through a linear layer and a weighted linear activation function in sequence; S14: Information flow 2 passes through another linear layer, a 3×3 depthwise separable convolutional layer, and a 2D selective scanning module. The process of the 2D selective scanning module includes three steps: S141: Directional expansion, for the input feature map of the two-dimensional selective scanning module , expanded into a one-dimensional sequence along four directions: （1）； in: Represents a function that expands a two-dimensional feature map into a one-dimensional sequence; Indicates the A sequence that expands in two directions; and Indicates scanning from upper left to lower right and from lower right to upper left in the horizontal direction; and Indicates scanning from upper left to lower right and from lower right to upper left in the vertical direction; S142: State space modeling, each sequence For causal sequence modeling, the discrete state space model is expressed as: （2）； in: Indicates the first The internal hidden state at each moment; represents the state transition matrix; represents the input transformation matrix; Indicates the first The input vector at each moment; represents the output features; represents the mapping matrix; and Discretization of state-space models from continuous form; （3）； in: and represents the original parameters of the continuous state space model; represents the time scale parameter used to map continuous parameters; represents the matrix exponential function; represents the identity matrix; S143: Reconstruction fusion, output sequence in each direction Reconstructed into feature map structure: （4）； in: Indicates restoring a one-dimensional sequence to a two-dimensional feature map by direction; Represents the restored two-dimensional feature map; Finally, the four directional outputs are fused into the output feature map of the two-dimensional selective scanning module: （5）； in, Represents element-wise addition fusion function; S15: Fusing the outputs of data stream 1 and data stream 2 by element-by-element multiplication; S16: The fused features pass through a linear layer to generate the output features of the Mamba block: （6）； in: and Represent the input and output of the Mamba block respectively; Represents a pixel-by-pixel multiplication operation; represents a linear layer; Represents the linear layer operation of data flow one through; Represents the linear layer operation of data flow two; represents the weighted linear activation function; represents the normalization layer; represents a two-dimensional selective scanning module; Represents a depthwise separable convolutional layer.

3. The cross-viewpoint geolocation method for UAVs based on a state-space model according to claim 1, characterized in that: The specific steps of S2 are: S21: Extract features from drone images using the state space feature extraction module of S1 , extracting features from satellite images , and are input into a linear layer and a depth-separable convolution layer with a convolution kernel of 3×3 respectively to extract the local semantic features of the drone and satellite local semantic features : （7）； in, and Represents the linear layers for processing drone and satellite data respectively; and Denote the depthwise separable convolutional layers for processing UAV and satellite data, respectively; S22: and Input into the cross-mamba block so that the cross-view features contain rich interaction information; The cross-mamba block process consists of two steps: S221: Obtain the state transfer matrix and transformation matrix by discretizing the state space model: ； （8）； in, and Transfer matrix and transformation matrix representing the discretization of the UAV; and Transfer matrix and transformation matrix representing satellite discretization; and Represents the original parameters of the UAV continuous state space model; and Represents the original parameters of the satellite continuous state space model; and represent the time scale parameters of UAV and satellite data, respectively; S222: Cross-perspective causal sequence modeling, and Expanded into one-dimensional sequences respectively, the output of the cross-mamba block is expressed as: （9）； in: and Represents the first The input vector at each moment; and represent the UAV and satellite features output by the cross-Mamba block, respectively; Indicates that the drone sequence is in The hidden state vector at the moment; Indicates that the satellite sequence is The hidden state vector at the moment; Represents the satellite observation matrix, which is used to Cross-view mapping into output features ; Represents the observation matrix of the UAV, which is used to Cross-view mapping into output features ; S23: and The features of the drone and satellite are fused by sending them to the normalization layer and the linear layer respectively, and performing element-wise multiplication: （10）； in: Represents the fused cross-view features; represents the linear layer used to process drone features; represents the linear layer used to process satellite features; represents the normalization layer used to process drone features; Represents a normalization layer used to process satellite features.

4. The cross-viewpoint geolocation method for UAVs based on a state-space model according to claim 1, characterized in that: The specific steps of S3 are: S31: For the input drone image and satellite imagery , first reduce their spatial resolution by half through bilinear interpolation method, then, the reduced resolution UAV images and satellite imagery Send to backbone network; S311: In the first stage of the backbone network, the features extracted by the twin feature extraction module are represented as: （11）； in: and They represent the Mamba blocks used for UAV and satellite image feature extraction in the first stage; and Represent the UAV and satellite features extracted in the first stage respectively; S312: The cross-view features extracted by the state space feature cross fusion module are expressed as: （12）； in: Represents the operation of the first-stage state space feature cross-fusion module; S313: The view-independent features extracted in the first stage and , as the input of the next stage, used to extract deep semantic features, Cross-view features fused in multiple stages Expressed as: （13）； in, and Respectively represent The Mamba block is used for feature extraction of UAV and satellite images; Indicates the Operation of the stage state space feature cross fusion module; and Respectively represent UAV and satellite features extracted in the first stage; S32: The cross-view features extracted in the four stages are input to the positioning neck. The obtained features are expressed as follows: （14）； in: Indicates a 2x upsampling operation; Indicates the A 1×1 convolutional layer; Indicates channel-based splicing operation; S33: Position the neck features Input to the positioning head, through the linear layer regression heat map, its resolution is the same as the satellite image Same, the maximum value in the heat map corresponds to the satellite image The coordinates of the image are the drone image The center point coordinates of the drone are obtained to achieve geographic positioning.

5. The cross-viewpoint geolocation method for UAVs based on a state-space model according to claim 1, characterized in that: The specific steps of S4 are: S41: Satellite imagery and drone imagery Perform data augmentation; S411: Expanding with random pruning strategies ,make and Satellite images The length and width, first of all, in satellite images Find drone imagery in Center coordinates; then, take the coordinates as the center and The random cropping length and width are and Partial image of ; Finally, in the local image Find a random coordinate and construct a structure with the length and width as the center. Augmented satellite imagery ; S412: Expanding with foreground mask strategy , first construct a Then, use the mask to map the drone image Finally, the pixels in the covered area are set to 0 to obtain the expanded drone image. ; S42: Model optimization; S421: Standard heat map supervision, original drone image and expanded satellite imagery As the input of the network in step S3, the heat map of the drone image is predicted , using the prediction heat map and the real heat map Calculate weight balance loss , as shown below: （15）； in: Represents the index of each pixel position in the image; Represents the weight factor for balancing positive and negative samples; Indicates the The actual thermal value corresponding to each pixel on the label; Indicates the corresponding The predicted thermal value of each pixel; S422: Contextual reasoning supervision of occluded areas, expanding satellite images and drone imagery As the network input in step S3, obtain the predicted heat map , using the prediction heat map and prediction heatmap Calculate mean squared error loss , as shown below: （16）； in: and Represent the length and width of the heat map respectively; S423: Jointly optimize the objective function, combine formulas (15) and (16), and calculate the joint loss function: （17）； in: represents the hyperparameter used to balance the weights of the two losses.

Images