Multi-terrain type unmanned aerial vehicle GNSS-R height inversion method

Abstract

This application discloses a GNSS-R altimeter inversion method for unmanned aerial vehicles (UAVs) with multiple surface types, including constructing a GNSS-R altimeter inversion model adapted to multiple surface types; determining the current surface type of the UAV, so as to determine a matching GNSS-R altimeter inversion model for the current surface type from the multi-surface-type UAV GNSS-R altimeter inversion model; and inputting the GNSS-R altimeter observations corresponding to the current surface type into the matched GNSS-R altimeter inversion model to determine the ground clearance of the UAV.

Description

Technical Field

[0001] This application relates to the field of automatic identification technology, and in particular to a GNSS-R altimeter inversion method for unmanned aerial vehicles with multiple surface types. Background Technology

[0002] In the field of altitude measurement technology for unmanned aerial vehicles (UAVs), Global Navigation Satellite System Reflected Signal (GNSS-R) technology is gradually becoming an important means of passive and accurate altitude measurement. GNSS-R-based UAV altitude measurement achieves altitude measurement by receiving reflected signals from satellites such as GPS and BeiDou through a receiver on the UAV without actively emitting energy.

[0003] The specific principle involves accurately measuring parameters such as the phase difference and time delay difference between the direct and reflected signals from the satellite, and combining this with information such as the satellite position, receiver, and ground reflection characteristics to calculate the aircraft's altitude. Constructing an altitude measurement inversion model is the core and crucial step in GNSS-R altimeter measurement for unmanned aerial vehicles.

[0004] However, current GNSS-R-based unmanned aerial vehicle altitude measurement has many technical shortcomings, which severely limit its application in different scenarios.

[0005] On the one hand, existing altimetry models suffer from severe simplification. Most existing altimetry inversion models are limited to specific environments and struggle to adapt to the altimetry requirements of various landform types.

[0006] On the other hand, in natural land environments, GNSS reflected signals are weak and the signal reflection path is relatively simple. Factors such as soil moisture on the reflecting surface and the presence of vegetation can significantly affect the observation of reflected signals, thus affecting the accuracy of height measurement.

[0007] In urban environments, tall buildings, glass curtain walls, and metal structures cause satellite signals to reflect multiple times between buildings, forming complex multipath signals. This results in multiple scattering peaks in the delay-Doppler diagram (DDM), which significantly reduces the accuracy of altitude measurements.

[0008] Therefore, how to construct a GNSS-R altimeter inversion model that is compatible with various land surface types in order to enable unmanned aerial vehicles to accurately calculate their altitude in different scenarios has become a technical bottleneck that urgently needs to be solved. Summary of the Invention

[0009] This application provides a technical solution for GNSS-R altimetry inversion of unmanned aerial vehicles with multiple surface types.

[0010] Specifically, a GNSS-R altimeter inversion method for multi-surface-type unmanned aerial vehicles includes:

[0011] Construct a GNSS-R altimeter inversion model for unmanned aerial vehicles that is compatible with multiple terrain types;

[0012] Determine the current surface type of the UAV so as to identify the UAV GNSS-R altimeter inversion model that matches the current surface type from the UAV GNSS-R altimeter inversion model with multiple surface types;

[0013] The GNSS-R altimeter measurements corresponding to the current surface type are input into the matched GNSS-R altimeter inversion model of the unmanned aerial vehicle (UAV) to determine the UAV's altitude above the ground.

[0014] The technical solution provided in this application has at least the following beneficial effects:

[0015] This application overcomes the limitation of a single altimeter model by constructing a GNSS-R altimeter inversion model for unmanned aerial vehicles (UAVs) that is adaptable to multiple surface types. This enables UAVs to have corresponding altimeter inversion models in different surface environments, such as land and urban blocks, thereby improving the applicability of the model and meeting the altimeter requirements of various surface types.

[0016] This application determines the current surface type of the unmanned aerial vehicle (UAV) and inputs the GNSS-R altimeter observations corresponding to the current surface type into the matching altimeter inversion model. Based on the characteristics of the natural land surface environment, the corresponding altimeter inversion model can be used to measure the altitude, reducing the influence of factors such as soil moisture and vegetation on the altimeter measurement and improving the accuracy of UAV altimeter measurement in natural land surface environments.

[0017] This application constructs an altimetry inversion model adapted to multiple land surface types. It can determine the altimetry inversion model that matches the characteristics of urban street environments. By inputting GNSS-R altimetry observations corresponding to the current land surface type into the model, it can effectively handle complex multipath signals in urban street environments, reduce the interference of multiple scattering peaks on altimetry, and improve the altimetry accuracy of unmanned aerial vehicles in urban street environments.

[0018] The technical solution of this application constructs a GNSS-R altimeter inversion model for unmanned aerial vehicles (UAVs) that is adapted to multiple surface types. Combined with the operation of determining the current surface type of the UAV and inputting the corresponding observations, the UAV can use a suitable altimeter inversion model to measure its altitude in various surface types and different scenarios. This achieves the goal of accurately calculating the ground clearance of the UAV in different scenarios and solves the technical bottleneck of the prior art that makes it difficult to accurately measure altitude in different scenarios. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1 A flowchart of the GNSS-R altimeter inversion method for multi-surface-type unmanned aerial vehicles provided for the implementation of this application.

[0021] Figure 2 A schematic diagram of a GNSS-R altimeter inversion model adapted for unmanned aerial vehicles on flat ground.

[0022] Figure 3 A schematic diagram of a GNSS-R altimeter inversion model adapted for unmanned aerial vehicles in mountainous terrain.

[0023] Figure 4 A schematic diagram of a GNSS-R altimeter inversion model adapted for urban unmanned aerial vehicles.

[0024] Figure 5 This is a schematic diagram illustrating the continuous training and optimization of the city model.

[0025] Figure 6 This is a schematic diagram of the CNN+LSTM model architecture. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] like Figure 1 As shown, a GNSS-R altimeter inversion method for unmanned aerial vehicles with multiple surface types includes:

[0028] Construct a GNSS-R altimeter inversion model for unmanned aerial vehicles that is compatible with multiple terrain types;

[0029] Determine the current surface type of the UAV so as to identify the UAV GNSS-R altimeter inversion model that matches the current surface type from the UAV GNSS-R altimeter inversion model with multiple surface types;

[0030] The GNSS-R altimeter measurements corresponding to the current surface type are input into the matched GNSS-R altimeter inversion model of the unmanned aerial vehicle (UAV) to determine the UAV's altitude above the ground.

[0031] Optionally, the multi-terrain-type UAV GNSS-R altimeter inversion model includes: an UAV GNSS-R altimeter inversion model adapted to flat land, mountainous land, and urban areas.

[0032] Optionally, a GNSS-R altimeter inversion model adapted for flat terrain is constructed, including:

[0033] Determine the flat terrain geometry and key flat terrain distances to be used in building the model;

[0034] Determine the flatland error factor for chip path delay to calculate the total flatland path delay;

[0035] The correlation between flat terrain geometry, key distances on flat terrain, total path delay on flat terrain, and altitude on flat terrain is determined in order to construct a GNSS-R altimeter inversion model for unmanned aerial vehicles adapted to flat terrain.

[0036] Optionally, the flat geometry includes the satellite elevation angle, the vector between the phase centers of the up-looking antenna and the down-looking antenna, and the flat key distance includes the altitude above the ground.

[0037] Optionally, the GNSS-R altimeter inversion model adapted for flat terrain is shown in the following formula: Where h represents the height above the ground. The total path delay difference between the direct signal from the satellite to the receiver and the reflected signal. The deviation between the direct signal channel and the reflected signal channel caused by hardware errors. The deviation of the reflective surface caused by topographic reflectivity and vegetation cover. Measurement error Unmodeled error Elevation angle of Global Navigation Satellite System (GNSS) satellites. The vector between the phase centers of the upward-looking antenna and the downward-looking antenna.

[0038] See Figure 2As shown, the technical essence of the GNSS-R altimeter inversion model adapted for flat ground is based on the principle of Global Navigation Satellite System reflected signals (GNSS-R). The UAV is equipped with specific antennas (an upward-facing right-hand circularly polarized antenna RHCP receives direct signals, and a downward-facing left-hand circularly polarized antenna LHCP receives reflected signals), and does not actively emit energy, only receiving direct and reflected signals from the satellite. By analyzing the parameter differences between these two signals, combined with relevant geometric relationships and error corrections, accurate measurement of the UAV's altitude above the ground is achieved, which belongs to passive altimeter technology. The underlying geometric relationships are as follows: As shown in Figure 1, there are several key geometric elements. Let the vertical distance from the phase center of the reflecting antenna to the ground be h, and the satellite elevation angle be... The vector between the phase centers of the upward-looking antenna and the downward-looking antenna is Assuming the direct and reflected signals are parallel (classic GNSS-R altimeter geometry), according to geometric principles, the path delay between the direct and reflected signals... Related to these geometric quantities, it can be deduced that This formula is based on the geometric relationships of right triangles, where 2h and The projections along the signal propagation direction together constitute the path delay-related geometric quantity. Comprehensive error considerations: In actual measurements, path delay is affected by various error interferences. Hardware deviations. Signal channel deviations are caused by inherent hardware differences; reflector deviations. Caused by topographic reflection characteristics (such as soil moisture and vegetation cover); measurement error Uncertainty arising from measurement process; unmodeled error This includes other factors that were not previously considered. Therefore, the actual total path delay difference... For geometric path delay The sum of all kinds of errors, i.e. The formula is derived as follows: Substitution After mathematical transformations (rearranging terms, simplification):

[0039]

[0040] This application integrates geometric relationships and actual errors, and can determine the total path delay difference based on the measurement. By using relevant error parameters, the ground clearance h of the unmanned aerial vehicle can be accurately calculated, and thus it can be used to express the GNSS-R altimeter inversion model of unmanned aerial vehicles adapted to flat terrain.

[0041] Optionally, a GNSS-R altimeter inversion model adapted for mountainous terrain is constructed, including:

[0042] Determine the mountain geometry and key mountain distances to be used in model building;

[0043] Determine the mountain error factor for chip path delay to calculate the total mountain path delay;

[0044] The correlation between mountain geometry, key distances, total path delay, and altitude was determined in order to construct a GNSS-R altimeter inversion model adapted to mountainous terrain.

[0045] Optionally, mountain geometry includes satellite elevation angle, mountain reflector slope angle, and the vector between the phase centers of the upward-looking antenna and the downward-looking antenna; key mountain distances include: altitude above ground.

[0046] Optionally, the GNSS-R altimeter inversion model adapted for mountainous terrain is shown in the following formula: Where h represents the height above the ground. The total path delay difference between the direct signal from the satellite to the receiver and the reflected signal. The deviation between the direct signal channel and the reflected signal channel caused by hardware errors. The deviation of the reflective surface caused by topographic reflectivity and vegetation cover. Measurement error Unmodeled error The vector between the phase centers of the upward-looking antenna and the downward-looking antenna. Elevation angle of Global Navigation Satellite System (GNSS) satellites. : The slope angle of the mountain's reflective surface.

[0047] See Figure 3 In this application, the GNSS-R altimeter inversion model adapted for mountainous terrain still essentially utilizes Global Navigation Satellite System (GNSS-R) reflected signals. The UAV receives direct signals via a right-hand circularly polarized antenna (RHCP) and reflected signals via a left-hand circularly polarized antenna (LHCP). Based on parameter analysis of these two signals (such as phase difference and time delay difference), combined with information such as satellite position, receiver, and mountain reflection characteristics, the altitude above the ground is calculated, which is a passive altimeter method.

[0048] In flat terrain ( Figure 2 In this case, the main consideration is the vertical distance from the phase center of the reflecting antenna to the ground. Satellite elevation angle Inter-antenna vector Etc. Path delay between direct and reflected signals. It is based on the simple geometric relationship of right triangles and does not take into account the terrain slope factor.

[0049] In mountainous scenes ( Figure 3 In this context, the ground has a slope, which increases the slope angle of the mountain's reflective surface. This is a crucial geometric element. The vertical distance from the phase center of the reflecting antenna to the reflecting surface is h', and the vertical distance to the ground is... ,and Path delay between direct and reflected signals .here This demonstrates the impact of slope on signal path delay calculation, which is fundamentally different from the flat terrain model. Because the slope of a mountain alters the geometric angle of the signal reflection path, the calculation of the signal propagation path must take into account the angular changes caused by the slope.

[0050] Optionally, construct an unmanned aerial vehicle GNSS-R altimeter inversion model adapted to the city, including:

[0051] Obtain the training data and label data for the constructed model;

[0052] The model training data is preprocessed to obtain the model training input data;

[0053] The model training input data is input into the model to be trained for forward propagation to obtain the predicted height value;

[0054] Based on the predicted height and the labeled data used for model training, calculate the loss function for model training;

[0055] Based on the loss function, backpropagation is performed on the model to be trained until the training of the model ends, resulting in an altimetry inversion model adapted to the city.

[0056] Optionally, the input data for model training includes two-dimensional DDM data, urban digital elevation data, time series data, and auxiliary data. The label data for model training is the measured true altitude value of the aircraft. The time series data includes SNR, carrier phase, and Doppler frequency shift data. The auxiliary data includes aircraft attitude angle, velocity, and acceleration data.

[0057] In this application, the GNSS-R altimeter inversion model for urban unmanned aerial vehicles (UAVs) is based on machine learning technology, primarily utilizing convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). Essentially, it extracts effective signal features by learning from and analyzing complex multipath satellite signal data in urban environments, thereby achieving accurate prediction of the UAV's altitude. Unlike flatland and mountainous models, which are based on geometric relationships and the physical principles of signal propagation, this model allows the machine to learn patterns from large amounts of data to perform altimeter measurements.

[0058] Compared to flatland and mountain models, which are primarily built upon geometric elements (such as satellite elevation angle and inter-antenna vectors) and the physical laws of signal propagation (considering path delay and error factors), flatland and mountain models rely on precise measurements and calculations of the signal's geometric path and related errors. The aircraft's altitude is then derived using geometric formulas. For example, flatland models are based on simple planar geometry, while mountain models consider the influence of terrain factors such as slope on these geometric relationships. In urban models, satellite signals in urban environments are affected by tall buildings, glass curtain walls, and metal structures, resulting in complex multipath effects. Figure 4 Traditional geometric models struggle to accurately depict this data. Therefore, models adapted to cities need to handle various data types. Input data includes two-dimensional DDM (Delay-Doppler Map) data, urban digital elevation data, time-series data (such as SNR, carrier phase, and Doppler shift), and auxiliary data (aircraft attitude angles, velocity, and acceleration). Preprocessing of this data (Figure 54), such as DDM data denoising, normalization, and image cropping, DEM data resampling, normalization, and registration, time-series signal standardization, and spatiotemporal synchronization, prepares the model for training. Then, CNN is used to automatically extract local spatial features from the DDM (such as scattering peak distribution and energy diffusion patterns), and LSTM is used to capture long-term dependencies in the time-series data (such as the continuous change trend of Doppler shift during rapid aircraft movement). Figure 6 (This is used to characterize the complex signal environment of a city by integrating spatial and temporal features.)

[0059] The process of building a city model is a continuous process of training and optimization. Figure 5 The preprocessed data is input into the CNN-LSTM model to be trained for forward propagation to obtain the height prediction value. A loss function is calculated based on the predicted value and the measured true height value (label data). The loss function comprehensively considers mean squared error (MSE) or Huber loss (to address data noise and outliers), terrain consistency loss (to ensure the predicted height matches the DEM elevation), and physical model constraint loss (embedding the physical laws of GNSS-R reflection). Then, the model parameters are continuously adjusted through backpropagation, while employing training strategies such as the Adam optimizer to accelerate convergence, regularization techniques (Dropout and L2 regularization) to prevent overfitting, early stopping strategies, and learning rate decay until the model converges, resulting in a height inversion model adapted to the city. This machine learning-based training and optimization process is fundamentally different from the construction methods of flatland and mountainous models.

[0060] See you again Figure 5 The technical approach for adapting GNSS-R altimeter inversion models for urban unmanned aerial vehicles is as follows:

[0061] 1. Data Preparation and Preprocessing

[0062] Data Acquisition: This involves collecting various data for model training. Input data includes 2D DDM (Delay-Doppler Map), city digital elevation model (DEM), time-series signals (such as SNR, carrier phase, and Doppler shift), and auxiliary signals (such as aircraft attitude angles, velocity, and acceleration). Label data consists of actual altitude values ​​obtained through lidar or ground-based measurements. This step corresponds to... Figure 4 The "Data Acquisition" module clarifies the data sources required for model training.

[0063] Data preprocessing: The collected data is processed. DDM data undergoes wavelet transform or median filtering for noise reduction, power values ​​are normalized to the [0,1] interval, and a 50×50 pixel area around the central peak is cropped and focused. DEM data is resampled to match the spatial resolution of the DDM, elevation values ​​are normalized to the [-1,1] interval, and then registered with the DDM image according to geographic coordinates. Time-series signals are segmented according to fixed time windows, and SNR and Doppler shift are Z-score normalized. Finally, the DDM, time-series signals, and auxiliary data are synchronized according to timestamps. These operations are performed in... Figure 4 The "Data Preprocessing" module aims to standardize the data format, remove noise interference, and provide high-quality data for subsequent model training.

[0064] Dataset partitioning: The preprocessed data is divided into training, validation, and test sets in a 7:2:1 ratio, while ensuring temporal continuity and preventing future data leaks. This is consistent with... Figure 4 Consistent with the "Dataset Partitioning" module, properly partitioning the dataset helps the model play different roles in the training, validation, and testing phases, and evaluates the model's generalization ability.

[0065] 2. Model Training and Optimization

[0066] Introducing a loss function: To balance data-driven learning, physical constraints, and terrain consistency requirements, a total loss function is constructed. Among them, the mean squared error (MSE) or Huber loss is selected as the main regression loss based on the data noise level; terrain consistency loss... This is used to constrain the predicted height to be consistent with the prior elevation of the DEM. Adjusted based on DEM confidence level; physical model constraint loss. This involves embedding the physical laws of GNSS-R reflection into the model to avoid overfitting. This part... Figure 4 The "Introducing Loss Function" module reflects this.

[0067] Data augmentation: Augmentation operations are performed on data with small sample sizes. DDM augmentation increases data diversity by randomly shifting and rotating locations and adding Gaussian noise; DEM augmentation simulates urban expansion by randomly adding virtual buildings or terrain undulations, improving the model's robustness to DEM errors; time-series signal augmentation simulates data at different scales by adding random time offsets and scaling perturbations, enhancing the model's generalization ability. Figure 4 The "Data Augmentation" module aims to expand the dataset and improve the model's generalization performance.

[0068] Pre-trained DEM branch and trained model: First, a DEM processing branch is pre-trained on a public DEM dataset to learn general terrain feature representations. Cross-city migration is then performed through DEM feature adaptation, and the entire model is trained afterwards. This is in... Figure 5 The “Pre-trained DEM Branch” and “Trained Model” modules demonstrate this. Pre-training can accelerate the training process, reduce data requirements, and improve model robustness.

[0069] 3. Model Validation and Deployment

[0070] Model validation: Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used to evaluate model performance. Ablation experiments were also conducted, removing the LSTM layer, CNN layer, and attention mechanism to verify the contribution of each component to model performance. This is consistent with... Figure 5 The corresponding "Model Validation" module evaluates the effectiveness of the model and the importance of each part through multi-dimensional assessment.

[0071] Model lightweighting: To improve deployment efficiency and reduce hardware costs, pruning is used to remove neurons or channels with low contribution, reducing model parameters. Quantization converts floating-point weights to 8-bit integers to reduce storage requirements. TensorRT is enabled for acceleration when deployed to edge devices. Figure 5 The “Model Lightweighting” module offers operations that can improve model inference speed and optimize resource utilization.

[0072] Deployment to the receiver: The validated and lightweight model is deployed to the receiver, enabling it to accurately measure the altitude of the unmanned aerial vehicle in real-time during practical applications. This is the final application stage of the entire process. Figure 5 "Deploy to Receiver" module.

[0073] See Figure 6 In the CNN+LSTM model architecture shown:

[0074] CNN (Convolutional Neural Network): Automatically extracts spatial features from data through structures such as convolutional layers and pooling layers.

[0075] LSTM (Long Short-Term Memory): This network processes time-series data and solves the gradient vanishing and gradient exploding problems that exist in traditional RNNs. It is good at capturing long-term dependencies in time series data.

[0076] DDM (Delay-Doppler Map): In GNSS-R (Global Navigation Satellite System Reflection Measurement), it is a graphical representation of the delay and Doppler shift information of satellite reflected signals, reflecting the characteristics of the signal during propagation.

[0077] DEM (Digital Elevation Model): A digital simulation of ground topography using limited terrain elevation data, expressing terrain elevation information in digital form.

[0078] Convolution: The core operation in CNNs, which uses a convolution kernel to slide across the data to perform weighted summation and extract local features from the data.

[0079] Pooling: Usually follows convolutional layers and is used to downsample feature maps, reducing the amount of data while retaining the main features. Common types include max pooling and average pooling.

[0080] Fully Connected Layer: In a neural network, all neurons in each layer are interconnected, used to integrate and map the features extracted earlier to achieve the final prediction.

[0081] Feature vector: A vector representation of the data features extracted after processing by each layer of a neural network, used to characterize the key information of the data.

[0082] See Figure 6 The principle behind the CNN+LSTM model architecture shown is explained below:

[0083] 1. Input Data Processing: Input data includes DDM data, DEM data, and time-series information containing SNR (Signal-to-Noise Ratio), Doppler shift, and aircraft motion status. First, these data undergo normalization, spatial registration, and time synchronization to ensure they meet model input requirements and remain consistent in time and space. Normalization adjusts the data to an appropriate range, spatial registration ensures accurate spatial correspondence between data from different sources, and time synchronization guarantees consistency of the time-series data across the time dimension.

[0084] 2. CNN branches extract spatial features

[0085] CNN Branch 1 (DDM Processing): DDM data enters CNN Branch 1. Through multiple convolutional operations, the convolutional kernels slide across the DDM image to extract features from local regions, such as capturing spatial features related to satellite reflection signals, like scattering peak distribution and energy diffusion patterns. After each convolutional layer, pooling operations (such as max pooling) are used to downsample the feature map, reducing data dimensionality while retaining important features. After multiple convolutions and pooling, global average pooling is finally used to compress the feature map into a 128-dimensional feature vector, effectively extracting and condensing the spatial features of the DDM data.

[0086] CNN Branch 2 (DEM Processing): DEM data is fed into CNN Branch 2. Considering the relatively static nature of DEM information, a shallower network structure is adopted. Through a small number of convolution and pooling operations, spatial feature information such as terrain elevation is extracted from the DEM data, and a 32-dimensional feature vector is finally output.

[0087] Feature concatenation: The 128-dimensional DDM feature vector output from CNN branch 1 and the 32-dimensional DEM feature vector output from CNN branch 2 are concatenated to form a 160-dimensional vector, which integrates the spatial features of DDM and DEM data and provides more comprehensive spatial information for subsequent processing.

[0088] 3. LSTM Branch Extracts Temporal Features: The input temporal signals include SNR, Doppler shift, flight attitude angles, etc., and DEM statistics, such as the current area's average elevation, are added. The final input dimension is T×F, where T is the time step and F is the number of features. A two-layer LSTM is used, with 128 units per layer, and the output for all time steps is returned. The LSTM layer, through its unique gating mechanism (input gate, forget gate, output gate), can remember and update information in the time series, effectively capturing the long-term dependencies between different time steps, such as the continuous change trend of Doppler shift over time during aircraft motion. Temporal attention is applied to the LSTM output, weighting and aggregating key time step features, and terrain-aware attention is added. The temporal weights are dynamically adjusted based on the DEM features; for example, if the DEM shows the current area is a flat road, the sensitivity to multipath interference signals is reduced. The final output of the LSTM branch is a 128-dimensional temporal feature vector.

[0089] 4. Feature Fusion and Regression Prediction

[0090] Feature fusion: In the fusion layer, the 160-dimensional vector output by the CNN and the 128-dimensional vector output by the LSTM are concatenated to generate a 288-dimensional vector, which realizes the fusion of spatial features and temporal features, enabling the model to comprehensively utilize the two types of feature information to describe the relevant characteristics of GNSS-R signals in urban street environments.

[0091] Regression Prediction: The fused 288-dimensional vector is input into the regression layer, which employs a 64-dimensional fully connected layer to integrate and map the input features. Dimensionality reduction of the features helps reduce computational cost, prevent overfitting, and extract more compact feature representations. The ReLU activation function is used to introduce non-linear features, enabling the model to learn more complex feature relationships. A single neuron with linear activation is used in the output layer of the regression layer to generate a continuous altitude prediction value, thus enabling the prediction of the unmanned aerial vehicle's altitude in urban street environments.

[0092] 5. Model Training and Optimization

[0093] Loss Function Construction: To balance data-driven learning, physical constraints, and terrain consistency requirements, the loss function consists of multiple parts. If the data noise is low and uniformly distributed, the mean squared error (MSE) is used for the main regression loss. When outliers exist (such as brief interruptions in GNSS signals), Huber loss is used to enhance robustness. Terrain consistency loss is also introduced. , The constrained predicted height is dynamically adjusted based on the DEM confidence level, ensuring consistency with the prior DEM elevation; a physical model constraint loss is introduced. The physical laws governing GNSS-R reflection (such as bistatic radar equations and geometric path delay) are embedded into the model to avoid overfitting driven by pure data. The total loss function is... .

[0094] Optimizer Selection and Parameter Tuning: The Adam optimizer was selected during training to leverage its adaptive properties and accelerate model convergence in complex multi-path signals. The initial learning rate was set to... In the initial stage, a larger learning rate is used to achieve rapid convergence, and in the later stage, the learning rate is reduced to refine the parameters.

[0095] Regularization to prevent overfitting: Dropout (ratio 0.3-0.5) is added after the fully connected layer to randomly discard neurons and introduce noise to enhance robustness; L2 regularization is used to apply L2 constraints (coefficient) to the weights of the convolutional and fully connected layers. (Adjusted according to model complexity and data volume), limiting the weight magnitude to control model complexity.

[0096] Training Strategy: Considering factors such as computational efficiency, the batch size is set to 32-64, adjusted according to GPU memory. An early stopping strategy is applied, monitoring the validation set loss; training is terminated if there is no improvement after 10 consecutive epochs to prevent overfitting, save resources, and dynamically adjust training. Learning rate decay is used, halving the learning rate during loss plateaus to improve accuracy and stability and avoid local optima. The DEM processing branch is pre-trained on a public DEM dataset to learn general terrain feature representations. Cross-city migration is then performed using DEM features to accelerate training, reduce data requirements, and improve model robustness. Data augmentation is performed for small samples. DDM augmentation increases diversity by randomly translating and rotating locations and adding Gaussian noise. DEM augmentation simulates city expansion by randomly adding virtual buildings or terrain undulations to enhance robustness to DEM errors. Time-series signal augmentation simulates data at different scales by adding random time offsets and scaling perturbations to improve the model's generalization ability and ability to cope with data fluctuations.

[0097] Corresponding to the above Figure 1 This application also provides a GNSS-R altimeter inversion device for multi-surface-type unmanned aerial vehicles, comprising:

[0098] The model building unit is used to build GNSS-R altimeter inversion models for unmanned aerial vehicles that are adapted to multiple terrain types.

[0099] The matching unit is used to determine the current surface type of the UAV so as to determine the UAV GNSS-R altimeter inversion model that matches the current surface type from the UAV GNSS-R altimeter inversion model with multiple surface types;

[0100] The measurement unit is used to input GNSS-R altimetry observations corresponding to the current surface type into the matched UAV GNSS-R altimetry inversion model to determine the UAV's altitude above the ground.

[0101] For an exemplary explanation of each of the above units, please refer to the above. Figure 1 .

[0102] Corresponding to the above Figure 1 This application also provides an electronic device including a memory and a processor. The memory stores a computer-executable program, and the processor runs the computer-executable program to implement the method of any of the above embodiments.

[0103] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A GNSS-R altimeter inversion method for unmanned aerial vehicles with multiple surface types, characterized in that, include: Construct a GNSS-R altimeter inversion model for unmanned aerial vehicles that is compatible with multiple terrain types; Determine the current surface type of the UAV so as to identify the UAV GNSS-R altimeter inversion model that matches the current surface type from the UAV GNSS-R altimeter inversion model with multiple surface types; Input the GNSS-R altimeter observations corresponding to the current surface type into the matched UAV GNSS-R altimeter inversion model to determine the UAV's altitude above the ground; Among them, the multi-surface-type UAV GNSS-R altimeter inversion model includes: UAV GNSS-R altimeter inversion model adapted to flat land, mountainous land and urban areas; The GNSS-R altimeter inversion model for urban unmanned aerial vehicles is a CNN-LSTM model, which includes: CNN Branch 1: Through multi-layer convolution operations, the convolution kernel slides on the DDM image to extract features from local regions, capturing spatial features related to scattering peak distribution, energy diffusion morphology and satellite reflection signals to obtain feature maps. After each convolution layer, pooling operation is used to downsample the feature maps, and finally global average pooling is used to obtain the DDM feature vector. CNN Branch 2: Extracts spatial features of terrain elevation from DEM data, outputs DEM feature vector and DDM feature vector and concatenates them to combine the spatial features of DDM image and DEM data; LSTM branch: The LSTM layer is used to capture the long-term dependencies between these time-series signals at different time steps through its gating mechanism to obtain the LSTM output. Temporal attention is applied to the LSTM output, key time-step features are weighted and aggregated, and terrain-aware attention is added. The temporal weights are dynamically adjusted according to the DEM features to finally output the temporal feature vector. Fusion layer: Concatenates the fusion vector of DEM feature vector and DDM feature vector with the temporal feature vector output by LSTM; Regression layer: The fused vector of DEM feature vector and DDM feature vector and the fused vector of temporal feature vector output by LSTM are input into the regression layer for integration and mapping. A single neuron linear activation is used in the output layer of the regression layer to generate a continuous altitude prediction value, thereby realizing the prediction of the ground clearance of the unmanned aerial vehicle in urban street environment.

2. The method according to claim 1, characterized in that, Constructing a GNSS-R altimeter inversion model for unmanned aerial vehicles adapted to flat terrain includes: Determine the flat terrain geometry and key flat terrain distances to be used in building the model; Determine the flatland error factor for chip path delay to calculate the total flatland path delay; Determine the relationship between flat terrain geometry, total path delay, and altitude on flat terrain to construct a GNSS-R altimeter inversion model for unmanned aerial vehicles adapted to flat terrain.

3. The method according to claim 2, characterized in that, The flat-ground geometry includes the satellite elevation angle, the vector between the phase centers of the up-looking antenna and the down-looking antenna, and the flat-ground critical distance includes the altitude above the ground.

4. The method according to claim 3, characterized in that, The GNSS-R altimeter inversion model for unmanned aerial vehicles adapted to flat terrain is shown in the following formula: Where h represents the height above the ground. The total path delay difference between the direct signal from the satellite to the receiver and the reflected signal. The deviation between the direct signal channel and the reflected signal channel caused by hardware errors. The deviation of the reflective surface caused by topographic reflectivity and vegetation cover. Measurement error Unmodeled error Elevation angle of Global Navigation Satellite System satellites The vector between the phase centers of the upward-looking antenna and the downward-looking antenna.

5. The method according to claim 1, characterized in that, Constructing a GNSS-R altimeter inversion model for unmanned aerial vehicles adapted to mountainous terrain includes: Determine the mountain geometry and key mountain distances to be used in model building; Determine the mountain error factor for chip path delay to calculate the total mountain path delay; The correlation between mountain geometry, key distances, total path delay, and altitude was determined in order to construct a GNSS-R altimeter inversion model adapted to mountainous terrain.

6. The method according to claim 5, characterized in that, Mountain geometry includes satellite elevation angle, mountain reflector slope angle, and the vector between the phase centers of the upward-looking antenna and the downward-looking antenna; Key distances in mountainous terrain include: altitude above ground.

7. The method according to claim 6, characterized in that, The GNSS-R altimeter inversion model for unmanned aerial vehicles adapted to mountainous terrain is shown in the following formula: Where h represents the height above the ground. The total path delay difference between the direct signal from the satellite to the receiver and the reflected signal. The deviation between the direct signal channel and the reflected signal channel caused by hardware errors. The deviation of the reflective surface caused by topographic reflectivity and vegetation cover. Measurement error Unmodeled error The vector between the phase centers of the upward-looking antenna and the downward-looking antenna. The elevation angle of satellites in a global navigation satellite system. : The slope angle of the mountain's reflective surface.

8. The method according to claim 1, characterized in that, Constructing an urban-adaptive GNSS-R altimeter inversion model for unmanned aerial vehicles, including: Obtain the training data and label data for the constructed model; The model training data is preprocessed to obtain the model training input data; The model training input data is input into the model to be trained for forward propagation to obtain the predicted height value; Based on the predicted height and the label data used for model training, the loss function for model training is calculated. Based on the loss function, backpropagation is performed on the model to be trained until the training of the model ends, resulting in an altimetry inversion model adapted to the city.

9. The method according to claim 8, characterized in that, The input data for model training includes two-dimensional DDM data, urban digital elevation data, time series data, and auxiliary data. The label data for model training is the actual measured altitude value of the aircraft. The time series data includes SNR, carrier phase, and Doppler frequency shift data. The auxiliary data includes aircraft attitude angle, velocity, and acceleration data.

Images