GNSS-R (Global Navigation Satellite System-Radio) altimetry inversion method for multi-surface type unmanned aerial vehicle

Abstract

The invention discloses a GNSS-R (Global Navigation Satellite System-Radio) altimetry inversion method for a multi-surface type unmanned aerial vehicle. The method comprises the following steps: constructing an unmanned aerial vehicle GNSS-R altimetry inversion model adaptive to multiple surface types; determining the current surface type of the unmanned aerial vehicle, so as to determine an unmanned aerial vehicle GNSS-R height measurement inversion model matched with the current surface type from the unmanned aerial vehicle GNSS-R height measurement inversion models of multiple surface types; and inputting the GNSS-R height measurement observed quantity corresponding to the current ground surface type into the matched unmanned aerial vehicle GNSS-R height measurement inversion model so as to measure the terrain clearance of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present application relates to the field of automatic identification technology, and in particular to a GNSS-R altimetry inversion method for unmanned aerial vehicles of multiple surface types. Background Art

[0002] In the field of unmanned aerial vehicle (UAV) altitude measurement, Global Navigation Satellite System Reflectometry (GNSS-R) technology has become an important means of passive precision altitude measurement. GNSS-R-based UAV altitude measurement relies on an onboard receiver to receive reflected signals from GPS, BeiDou, and other satellites without actively transmitting energy.

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

[0004] However, the current UAV altitude measurement based on GNSS-R technology has many technical defects, which seriously limit its application in different scenarios.

[0005] On the one hand, existing altimetry models are severely limited to specific environments and are difficult to adapt to the altimetry needs of various surface types.

[0006] On the other hand, in a natural land surface environment, the GNSS reflected signal intensity is weak and the signal reflection path is relatively single. Factors such as the soil moisture of the reflecting surface and the presence of vegetation will have a significant impact on the observed amount of the reflected signal, thereby affecting the accuracy of the height measurement.

[0007] In urban block environments, high-rise buildings, glass curtain walls, and metal structures can cause satellite signals to reflect multiple times between buildings, forming complex multipath signals and multiple scattering peaks in the Delay-Doppler Map (DDM), which greatly reduces the accuracy of altitude measurement.

[0008] Therefore, how to build a GNSS-R altimetry inversion model that is suitable for various surface types to enable unmanned aerial vehicles to accurately calculate their height above the ground in different scenarios has become a technical bottleneck that needs to be solved urgently. Summary of the Invention

[0009] The embodiment of the present application provides a technical solution for recognizing smoking behavior with high accuracy.

[0010] Specifically, a multi-surface type unmanned aerial vehicle GNSS-R altimetry inversion method includes:

[0011] Construct an unmanned aerial vehicle GNSS-R altimetry inversion model that is suitable for multiple surface types;

[0012] Determining the type of the surface the UAV is currently located on, so as to determine a UAV GNSS-R altimetry inversion model that matches the current surface type from among the UAV GNSS-R altimetry inversion models of multiple surface types;

[0013] The GNSS-R altimetry observations corresponding to the current surface type are input into the matched UAV GNSS-R altimetry inversion model to determine the UAV's height above the ground.

[0014] The technical solutions provided in the embodiments of the present application have at least the following beneficial effects:

[0015] This application breaks through the limitation of the single altimetry model by constructing a GNSS-R altimetry inversion model for unmanned aerial vehicles that is suitable for multiple surface types. It enables unmanned aerial vehicles to have corresponding altimetry inversion models in different surface types, such as land, urban blocks, etc., improves the applicability of the model and meets the altimetry needs of various surface types.

[0016] In this application, the current surface type of the unmanned aerial vehicle is determined, and the GNSS-R altimetry observations corresponding to the current surface type are input into the matching altimetry inversion model. According to the characteristics of the natural land surface environment, the corresponding altimetry inversion model can be used to measure altitude, reducing the influence of factors such as soil moisture and vegetation on altimetry, and improving the accuracy of altimetry of unmanned aerial vehicles in natural land surface environments.

[0017] This application constructs an altimetry inversion model that is adaptable to multiple surface types. It can determine an altimetry inversion model that matches the characteristics of the urban block environment, and input the GNSS-R altimetry observations corresponding to the current surface type into the model. It can effectively process the complex multipath signals in the urban block environment, reduce the interference of multiple scattering peaks on the altimetry, and improve the altimetry accuracy of unmanned aerial vehicles in the urban block environment.

[0018] The technical solution of the present application constructs a GNSS-R altimetry inversion model for unmanned aerial vehicles that is adaptable to multiple surface types, and combines it with the operation of determining the current surface type of the unmanned aerial vehicle and inputting the corresponding observation quantity, so that the unmanned aerial vehicle can use a suitable altimetry inversion model to measure altitude under various surface types and different scenarios, thereby achieving the goal of accurately calculating the height of the unmanned aerial vehicle above the ground in different scenarios, and solving the technical bottleneck of the existing technology that makes it difficult to accurately measure altitude in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0020] Figure 1 Flowchart of the multi-surface type UAV GNSS-R altimetry inversion method provided for the implementation of this application.

[0021] Figure 2 Schematic diagram of the UAV GNSS-R altimetry inversion model adapted for flat terrain.

[0022] Figure 3 Schematic diagram of the UAV GNSS-R altimetry inversion model adapted for mountainous areas.

[0023] Figure 4 Schematic diagram of the UAV GNSS-R altimetry inversion model adapted for cities.

[0024] Figure 5 A schematic diagram of continuous training and optimization of the city model.

[0025] Figure 6 Schematic diagram of the CNN+LSTM model architecture. DETAILED DESCRIPTION

[0026] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0027] like Figure 1 As shown, a multi-surface type unmanned aerial vehicle GNSS-R altimetry inversion method includes:

[0028] Construct an unmanned aerial vehicle GNSS-R altimetry inversion model that is suitable for multiple surface types;

[0029] Determining the type of the surface the UAV is currently located on, so as to determine a UAV GNSS-R altimetry inversion model that matches the current surface type from among the UAV GNSS-R altimetry inversion models of multiple surface types;

[0030] The GNSS-R altimetry observations corresponding to the current surface type are input into the matched UAV GNSS-R altimetry inversion model to determine the UAV's height above the ground.

[0031] Optionally, the multi-surface type UAV GNSS-R altimetry inversion model includes: an UAV GNSS-R altimetry inversion model adapted for flat land, mountainous area and urban area.

[0032] Optionally, a UAV GNSS-R altimetry inversion model suitable for flat terrain is constructed, including:

[0033] Determine the flat ground geometric elements and flat ground key distances used to build the model;

[0034] Determine the flat-earth error factor of the chip path delay to calculate the flat-earth total path delay;

[0035] The correlation between flat ground geometric elements, flat ground critical distance, flat ground total path delay and ground height is determined to construct an unmanned aerial vehicle GNSS-R altimetry inversion model suitable for flat ground.

[0036] Optionally, the flat-earth geometric elements include a satellite elevation angle, a vector between phase centers of an upward-looking antenna and a downward-looking antenna, and the flat-earth critical distance includes height above the ground.

[0037] Alternatively, the UAV GNSS-R altimetry inversion model adapted for flat terrain is as shown in the following formula:

[0038] Where h is the height above the ground, ΔD is the total path delay difference between the direct signal and the reflected signal from the satellite to the receiver, and δ hw : Deviation between the direct signal path and the reflected signal path caused by hardware error, δ r : The reflection surface deviation caused by terrain reflection characteristics and vegetation cover, δ m : measurement error, ε: unmodeled error, θ: elevation angle of the Global Navigation Satellite System (GNSS) satellite, The vector between the phase centers of the antenna looking up and looking down.

[0039] See also Figure 2 As shown in the figure, the technical essence of the GNSS-R altitude inversion model for UAVs adapted to flat land is based on the principle of global navigation satellite system reflected signals (GNSS-R). The UAV is equipped with specific antennas (the upward right-hand circularly polarized antenna RHCP receives direct signals, and the downward left-hand circularly polarized antenna LHCP receives reflected signals). It does not actively transmit energy, but only receives direct and reflected signals from satellites. By analyzing the parameter differences between the two signals, combining relevant geometric relationships and error correction, the accurate measurement of the aircraft's height above the ground is achieved, which is a passive altitude measurement technology. The geometric relationship basis of the expression reason: from Figure 1As can be seen from the figure, there are several key geometric elements. Let the vertical distance from the phase center of the reflector antenna to the ground be h, the satellite elevation angle be θ, and the vector between the phase centers of the upward-looking antenna and the downward-looking antenna be Assuming that the direct and reflected signals are parallel (classic GNSS-R altimetry geometry), according to geometric principles, the path delay Δρ between the direct and reflected signals is related to these geometric quantities, and it can be deduced that This formula is based on the geometric relationship of a right triangle, 2h and The projection in the signal propagation direction together constitutes the path delay related geometric quantity. Comprehensive error consideration: In actual measurement, path delay is affected by multiple errors. Hardware deviation δ hw Signal channel deviation caused by hardware differences; reflection surface deviation δ r Caused by terrain reflectance characteristics (such as soil moisture, vegetation cover); measurement error δ m The uncertainty comes from the measurement process; the unmodeled error ε covers other factors that are not taken into account. Therefore, the actual total path delay difference ΔD is the sum of the geometric path delay Δρ and various errors, that is, ΔD = Δρ + δ hw +δ r +δ m +ε. The formula is derived as follows: Substitute ΔD = Δρ + δ hw +δ r +δ m +ε, after mathematical transformation (transfer, simplification):

[0040]

[0041] This application combines geometric relationships with actual errors, and can accurately calculate the UAV height above the ground h based on the measured total path delay difference ΔD and related error parameters. Therefore, it can be used to express a UAV GNSS-R height measurement inversion model adapted for flat ground.

[0042] Optionally, a UAV GNSS-R altimetry inversion model suitable for mountainous areas is constructed, including:

[0043] Determine the mountain geometry elements and mountain key distances used to construct the model;

[0044] Determine the mountain error factor of the chip path delay to calculate the total mountain path delay;

[0045] The correlation between mountain geometric elements, mountain critical distance, mountain total path delay and altitude is determined to construct an unmanned aerial vehicle GNSS-R altitude inversion model suitable for mountains.

[0046] Optionally, the mountain geometric elements include the satellite elevation angle, the slope angle of the mountain reflector surface, and the vector between the phase centers of the upward-looking antenna and the downward-looking antenna; the mountain key distance includes: height above the ground.

[0047] Alternatively, the GNSS-R altimetry inversion model for unmanned aerial vehicles adapted to mountainous terrain is as shown in the following formula:

[0048] Where, h is the height above the ground, ΔD is the total path delay difference between the direct signal and the reflected signal from the satellite to the receiver, and δ hw : Deviation between the direct signal path and the reflected signal path caused by hardware error, δ r : the reflection surface deviation caused by terrain reflection characteristics and vegetation cover, δ m : measurement error, ε: unmodeled error, The vector between the phase centers of the antenna looking up and the antenna looking down, θ: the elevation angle of the Global Navigation Satellite System (GNSS) satellite, The slope angle of the mountain reflecting surface.

[0049] See also Figure 3 In this application, the GNSS-R altimetry inversion model for mountain-adapted UAVs still utilizes the Global Navigation Satellite System (GNSS-R) reflected signal technology. The UAV receives direct signals through its onboard right-hand circularly polarized antenna (RHCP) and reflected signals through its onboard left-hand circularly polarized antenna (LHCP). The altitude above the ground is calculated based on parameter analysis of these two signals (such as phase difference and time delay), combined with information such as satellite position, receiver, and mountain reflection characteristics. This is a passive altimetry method.

[0050] In the flatland scene ( Figure 2 ), the main considerations are the vertical distance h from the phase center of the reflecting antenna to the ground, the satellite elevation angle θ, and the vector between the antennas. etc. Direct signal and reflected signal path delay It is based on the geometric relationship of a simple right triangle and does not involve terrain slope factors.

[0051] In the mountain scene ( Figure 3 ) in the figure, the ground has a slope, which increases the slope angle of the mountain reflection surface. This is a key geometric element. The vertical distance from the phase center of the reflector antenna to the reflecting surface is h', and the vertical distance from the phase center to the ground is h, and Path delay between direct and reflected signals here This demonstrates the impact of slope on signal path delay calculations, which is fundamentally different from the flat-land model. Because mountain slopes change the geometric angle of the signal reflection path, the calculation of the signal propagation path must account for the angle changes caused by the slope.

[0052] Optionally, a UAV GNSS-R altimetry inversion model suitable for the city is constructed, including:

[0053] Obtain the training data and label data for the constructed model;

[0054] Preprocessing the model training data to obtain model training input data;

[0055] Inputting the model training input data into the model to be trained for forward propagation to obtain a measured height prediction value;

[0056] Calculating a loss function for model training based on the measured height prediction value and the label data for model training;

[0057] Based on the loss function, backpropagation is performed on the training model until the training of the model is completed, and a height measurement inversion model suitable for the city is obtained.

[0058] 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 actual height value of the aircraft obtained by actual measurement. The time series data includes SNR, carrier phase, and Doppler frequency shift data. The auxiliary data includes aircraft attitude angle, velocity, and acceleration data.

[0059] In this application, the city-adapted UAV GNSS-R altimetry inversion model is based on machine learning technology, mainly using convolutional neural networks (CNN) and long short-term memory networks (LSTM). Its essence is to extract effective signal features by learning and analyzing the relevant data of complex multi-path satellite signals in urban environments, and to achieve accurate prediction of the UAV's altitude above the ground. Unlike the flat and mountain models 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 altitude measurement.

[0060] Compared with the difference between flat land and mountain models, the flat land and mountain models are mainly constructed based on geometric elements (such as satellite elevation angle, inter-antenna vector, etc.) and the physical laws of signal propagation (taking into account path delay, error factors, etc.). By accurately measuring and calculating the geometric path of the signal and related errors, the height of the aircraft above the ground is derived using geometric formulas. For example, the flat land model is based on simple plane geometric relationships, and the mountain model considers the influence of terrain factors such as slope on the geometric relationship on this basis. In the urban model, satellite signals in the urban environment are affected by high-rise buildings, glass curtain walls and metal structures, and there are complex multipath effects ( Figure 4), traditional geometric models are difficult to accurately describe. Therefore, the model adapted to the city needs to process multiple types of data. The input data includes two-dimensional DDM (delay-Doppler map) data, urban digital elevation data, time series data (such as SNR, carrier phase, Doppler shift) and auxiliary data (aircraft attitude angle, velocity, acceleration). By preprocessing these data ( Figure 5 ), such as DDM data denoising, normalization, image cropping, DEM data resampling, normalization and registration, time series signal standardization and time-space synchronization, etc., to prepare for model training. Then CNN is used to automatically extract local spatial features in DDM (such as scattering peak distribution and energy diffusion morphology), and LSTM captures the long-term dependencies of time series data (such as the continuous change trend of Doppler frequency shift when the aircraft moves rapidly). Figure 6 ), integrating spatial and temporal features to characterize the complex signal environment in the city.

[0061] The process of building a city model is a process of continuous 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. The loss function is calculated based on the predicted value and the measured true height value (label data). The loss function comprehensively considers the mean square error (MSE) or Huber loss (to deal with data noise and outliers), terrain consistency loss (to constrain the predicted height to be consistent with the DEM elevation), and physical model constraint loss (to embed the physical law of GNSS-R reflection). The model parameters are then continuously adjusted through backpropagation. At the same time, the Adam optimizer is used to accelerate convergence, regularization techniques (Dropout and L2 regularization) are used to prevent overfitting, early stopping strategies and learning rate decay are used, and other training strategies are used until the model converges to obtain a height measurement inversion model suitable for the city. This machine learning-based training optimization process is essentially different from the way flat and mountain models are constructed.

[0062] See again Figure 5 The technical route of the GNSS-R altimetry inversion model for UAVs adapted to cities is as follows:

[0063] 1. Data preparation and preprocessing

[0064] Data collection: Collect various data for model training. Input data includes two-dimensional DDM (delay-Doppler map), urban digital elevation model (DEM), time series signals (such as SNR, carrier phase, Doppler shift) and auxiliary signals (such as aircraft attitude angle, speed, acceleration). Label data is the real height value obtained by lidar or ground measurement. This step corresponds to Figure 4 The "Data Collection" module clarifies the data sources required for model training.

[0065] Data preprocessing: The collected data is processed. DDM data is denoised by wavelet transform or median filtering, the power value is normalized to the [0,1] interval, and the area of 50×50 pixels around the central peak is cropped and focused. DEM data is resampled to make it the same as the DDM spatial resolution, the elevation value is normalized to the [-1,1] interval, and then aligned with the DDM image according to geographic coordinates. The time series signal is segmented according to a fixed time window, and the SNR and Doppler shift are Z-score standardized. Finally, the DDM, time series signal, and auxiliary data are synchronized according to the timestamp. These operations are performed in Figure 4 The "data preprocessing" module in the dataset is used to unify the data format, remove noise interference, and provide high-quality data for subsequent model training.

[0066] Dataset division: The pre-processed data is divided into training set, validation set, and test set in a ratio of 7:2:1, while ensuring time continuity to prevent future data leakage. Figure 4 The "Dataset Division" module in the ,reasonable division of the dataset helps the model play different roles in the training, validation and testing stages, and evaluates the model's generalization ability.

[0067] 2. Model training and optimization

[0068] Introducing loss function: To balance data-driven learning, physical law constraints and terrain consistency requirements, a total loss function L is constructed total =L MSE / huber +L terrain +L physical Among them, the mean square error (MSE) or Huber loss is selected as the main regression loss according to the data noise situation; the terrain consistency loss

[0069] L terrain =λ terrain ·MSE(H pred ,H DEM ) is used to constrain the predicted height to be consistent with the DEM prior elevation, λ terrain Adjusted according to DEM confidence; physical model constraint loss L physical =λ physical ·MSE(H pred ,H physical ), embed the GNSS-R reflection physical law into the model to avoid overfitting. Figure 4 This is reflected in the "Introduce Loss Function" module.

[0070] Data enhancement: In the case of small samples, data enhancement operations are performed. DDM enhancement increases data diversity by randomly translating and rotating the position and adding Gaussian noise; DEM enhancement simulates urban expansion by randomly adding virtual buildings or terrain undulations to improve the model's robustness to DEM errors; time series signal enhancement simulates data of different scales by adding random time offsets and scaling perturbations to enhance the model's generalization ability. Figure 4 The purpose of the "data enhancement" module is to expand the data set and improve the generalization performance of the model.

[0071] Pre-training DEM branch and training model: First pre-train the DEM processing branch on the public DEM dataset to learn the general terrain feature expression, perform cross-city migration through DEM feature adaptation, and then train the entire model. Figure 5 The "Pre-trained DEM branch" and "Training model" modules are demonstrated. Pre-training can accelerate the training process, reduce data requirements, and improve model robustness.

[0072] 3. Model Validation and Deployment

[0073] Model validation: Root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R 2 ) to evaluate the model performance, and also through ablation experiments, remove the LSTM layer, CNN layer, and attention mechanism respectively to verify the contribution of each component to the model performance. Figure 5 Corresponding to the "Model Verification" module in the model, the effectiveness of the model and the importance of each part are verified through multi-dimensional evaluation.

[0074] Model lightweighting: To improve deployment efficiency and reduce hardware costs, pruning is used to remove neurons or channels with small contributions to reduce model parameters, and floating-point weights are converted to 8-bit integers to reduce storage requirements. TensorRT acceleration is enabled when deployed to edge devices. Figure 5 "Model Lightweight" module, these operations can improve model inference speed and optimize resource utilization.

[0075] Deploy to receiver: Deploy the verified and lightweight model to the receiver, so that it can accurately measure the altitude of the UAV in real time in actual applications. This is the final application link of the entire process, corresponding to Figure 5 Deploy to Receiver module.

[0076] See also Figure 6 In the CNN+LSTM model architecture shown:

[0077] CNN (Convolutional Neural Network): Automatically extracts spatial features of data through structures such as convolutional layers and pooling layers.

[0078] LSTM (Long Short-Term Memory): It processes time series data, solves the gradient vanishing and gradient exploding problems in traditional RNNs, and excels at capturing long-term dependencies in time series.

[0079] Delay-Doppler Map (DDM): In GNSS-R (Global Navigation Satellite System Reflectometry), a graphical representation of the delay and Doppler shift of satellite-reflected signals, reflecting the characteristics of the signal during propagation.

[0080] DEM (Digital Elevation Model): A digital simulation of ground terrain using limited terrain elevation data, expressing terrain elevation information in digital form.

[0081] Convolution: The core operation in CNN, which extracts local features of the data by sliding the convolution kernel over the data and performing weighted summation.

[0082] Pooling: Usually follows the convolutional layer and is used to downsample the feature map to reduce the amount of data while retaining the main features. Common pooling methods include maximum pooling and average pooling.

[0083] Fully Connected Layer: In a neural network, all neurons between layers are interconnected, which is used to integrate and map the previously extracted features to achieve the final prediction.

[0084] Feature vector: A vector representation of the data features extracted after processing by each layer of the neural network, used to represent the key information of the data.

[0085] See also Figure 6 The CNN+LSTM model architecture shown in the figure implements the following model construction principles:

[0086] 1. Input Data Processing: Input data includes DDM data, DEM data, and time series information including signal-to-noise ratio (SNR), Doppler shift, and aircraft motion status. This data is first normalized, spatially aligned, and time-synchronized to ensure it meets model input requirements and remains consistent across time and space. Normalization adjusts the data to an appropriate range, spatial alignment ensures accurate spatial alignment of data from different sources, and time synchronization ensures consistency across the time dimension of the time series data.

[0087] 2. CNN branch extracts spatial features

[0088] CNN branch 1 (DDM processing): DDM data enters CNN branch 1. Through multi-layer convolution operations, the convolution kernel slides across the DDM image, extracting features from local regions. For example, it captures spatial features related to satellite reflection signals, such as the distribution of scattering peaks and energy diffusion patterns. After each convolution layer, pooling operations (such as max pooling) are used to downsample the feature map to reduce the data dimension while retaining important features. After multiple layers of convolution and pooling, global average pooling is finally used to compress the feature map into a 128-dimensional feature vector, effectively extracting and concentrating the spatial features of the DDM data.

[0089] CNN Branch 2 (DEM Processing): DEM data enters CNN Branch 2. Considering the relative static nature of DEM information, a shallower network structure is employed. Through a small number of convolution and pooling operations, spatial features such as terrain elevation are extracted from the DEM data, ultimately outputting a 32-dimensional feature vector.

[0090] Feature splicing: The 128-dimensional DDM feature vector output by CNN branch 1 and the 32-dimensional DEM feature vector output by CNN branch 2 are spliced 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.

[0091] 3. LSTM branch extracts time series features: The input time series signal includes signal-to-noise ratio (SNR), Doppler shift, flight attitude angle, and other features. DEM statistics, such as the average elevation of the current area, are also added. The final input dimension is T × F, where T is the time step and F is the number of features. Two LSTM layers, each with 128 units, are used and return outputs for all time steps. Through its unique gating mechanism (input gate, forget gate, and output gate), the LSTM layer memorizes and updates information in the time series, effectively capturing long-term dependencies between time steps, such as the continuous temporal trend of Doppler shift during aircraft motion. Temporal attention is applied to the LSTM output, weighting and aggregating key time step features. Terrain-aware attention is also added to dynamically adjust the time series weights based on DEM features. For example, if the DEM indicates a flat road in the current area, sensitivity to multipath interference signals is reduced. The LSTM branch ultimately outputs a 128-dimensional time series feature vector.

[0092] 4. Feature fusion and regression prediction

[0093] 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. This achieves the fusion of spatial and temporal features, enabling the model to comprehensively utilize these two types of feature information to describe the relevant characteristics of GNSS-R signals in urban block environments.

[0094] Regression Prediction: The fused 288-dimensional vector is input to the regression layer, which uses a 64-dimensional fully connected layer to integrate and map the input features. This dimensionality reduction helps reduce computational effort, prevent overfitting, and extract a more compact feature representation. The ReLU activation function introduces nonlinear features, enabling the model to learn more complex feature relationships. The output layer of the regression layer uses single-neuron linear activation to generate a continuous altitude prediction value, thereby predicting the UAV's ground clearance in urban environments.

[0095] 5. Model training and optimization

[0096] Loss function construction: To balance data-driven learning, physical constraints, and terrain consistency requirements, the loss function consists of multiple parts. If the data has little noise and is evenly distributed, the main regression loss uses the mean square error (MSE). When there are outliers (such as a brief interruption of the GNSS signal), the Huber loss is used to enhance robustness. The terrain consistency loss L is introduced. terrain =λ terrain ·MSE(H pred ,H DEM ), λ terrain Dynamically adjust according to DEM confidence, constrain the predicted height to be consistent with the DEM prior elevation; introduce the physical model constraint loss L physical =λ physical ·MSE(H pred ,H physical ), embed the GNSS-R reflection physical laws (such as bistatic radar equations and geometric path delay) into the model to avoid pure data-driven overfitting. The total loss function is L total =L MSE / huber +L terrain +L physical .

[0097] Optimizer Selection and Parameter Adjustment: The Adam optimizer was used during training, leveraging its adaptive nature to accelerate model convergence in complex multipath signals. The initial learning rate was set to 1e-4. A higher learning rate was used initially for rapid convergence, and later, the learning rate was reduced to refine the parameters.

[0098] Regularization to prevent overfitting: Dropout (ratio 0.3-0.5) is added after the fully connected layer to randomly discard neurons to introduce noise and enhance robustness; L2 regularization is used to impose L2 constraints on the weights of the convolutional and fully connected layers (coefficient λ = 1e-4, adjusted according to model complexity and data volume) to limit the weight amplitude and control model complexity.

[0099] Training Strategy: Taking into account factors such as computational efficiency, the batch size is set to 32-64 and adjusted based on GPU memory. An early stopping strategy is applied, monitoring the validation set loss and terminating training after 10 consecutive epochs without improvement to prevent overfitting, conserve resources, and dynamically adjust training. Learning rate decay is employed, halving the learning rate when the loss plateaus, improving accuracy and stability while avoiding local optima. The DEM processing branch is pre-trained on a public DEM dataset to learn common terrain feature representations. DEM feature adaptation is then used for cross-city migration, accelerating training, reducing data requirements, and improving model robustness. Data augmentation is performed when the sample size is small. DDM augmentation increases diversity by randomly shifting and rotating the positions and adding Gaussian noise. DEM augmentation simulates urban expansion by randomly adding virtual buildings or terrain fluctuations, enhancing robustness to DEM errors. Time series signal augmentation simulates data at different scales by adding random time offsets and scaling perturbations, improving model generalization and its ability to cope with data fluctuations.

[0100] Corresponding to the above Figure 1 The present application also provides a GNSS-R altimetry inversion device for unmanned aerial vehicles of multiple surface types, comprising:

[0101] A model building unit, used to build an unmanned aerial vehicle GNSS-R altimetry inversion model that is suitable for multiple surface types;

[0102] a matching unit, configured to determine the type of surface the UAV is currently located on, so as to determine a UAV GNSS-R altimetry inversion model that matches the current surface type from among the UAV GNSS-R altimetry inversion models of multiple surface types;

[0103] The measurement unit is used to input the GNSS-R altimetry observation corresponding to the current surface type into the matched unmanned aerial vehicle GNSS-R altimetry inversion model to measure the height of the unmanned aerial vehicle above the ground.

[0104] For an exemplary explanation of each of the above units, please refer to the above Figure 1 .

[0105] Corresponding to the above Figure 1 An embodiment of the present application further provides an electronic device, which includes a memory and a processor, wherein a computer executable program is stored in the memory, and the processor runs the computer executable program to implement the method of any of the above embodiments.

[0106] It should be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, the phrase "comprises a..." to define an element does not preclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0107] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A multi-surface type unmanned aerial vehicle GNSS-R height measurement inversion method, characterized in that: include: Construct an unmanned aerial vehicle GNSS-R altimetry inversion model that is suitable for multiple surface types; Determining the type of the surface the UAV is currently located on, so as to determine a UAV GNSS-R altimetry inversion model that matches the current surface type from among the UAV GNSS-R altimetry inversion models of multiple surface types; The GNSS-R altimetry observations corresponding to the current surface type are input into the matched UAV GNSS-R altimetry inversion model to determine the UAV's height above the ground.

2. The method according to claim 1, characterized in that The multi-surface type unmanned aerial vehicle GNSS-R altimetry inversion model includes: an unmanned aerial vehicle GNSS-R altimetry inversion model adapted for flat land, mountainous areas, and cities.

3. The method according to claim 2, characterized in that Construct a UAV GNSS-R altimetry inversion model suitable for flat terrain, including: Determine the flat ground geometric elements and flat ground key distances used to build the model; Determine the flat-earth error factor of the chip path delay to calculate the flat-earth total path delay; The correlation between the total path delay of flat ground geometric elements and the height above the ground is determined to construct an unmanned aerial vehicle GNSS-R altimetry inversion model suitable for flat ground.

4. The method according to claim 3, characterized in that Flat-Earth geometry includes the satellite elevation angle, the vector between the phase centers of the upward-looking antenna and the downward-looking antenna, and flat-Earth critical distances include the height above the ground.

5. The method according to claim 4, characterized in that The GNSS-R altimetry inversion model for unmanned aerial vehicles adapted to flat terrain is shown in the following formula: Where h is the height above the ground, ΔD is the total path delay difference between the direct signal and the reflected signal from the satellite to the receiver, and δ hw : Deviation between the direct signal path and the reflected signal path caused by hardware error, δ r : The reflection surface deviation caused by terrain reflection characteristics and vegetation cover, δ m : measurement error, ε: unmodeled error, θ: elevation angle of the GNSS satellite, The vector between the phase centers of the antenna looking up and looking down.

6. The method according to claim 2, characterized in that Build a UAV GNSS-R altimetry inversion model suitable for mountainous areas, including: Determine the mountain geometry elements and mountain key distances used to construct the model; Determine the mountain error factor of the chip path delay to calculate the total mountain path delay; The correlation between mountain geometric elements, mountain critical distance, mountain total path delay and altitude above the ground is determined to construct an unmanned aerial vehicle GNSS-R altitude inversion model suitable for mountainous areas.

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

8. The method according to claim 7, characterized in that The GNSS-R altimetry inversion model for unmanned aerial vehicles adapted to mountainous terrain is shown in the following formula: Where h is the height above the ground, ΔD is the total path delay difference between the direct signal and the reflected signal from the satellite to the receiver, and δ hw : Deviation between the direct signal path and the reflected signal path caused by hardware error, δ r : the reflection surface deviation caused by terrain reflection characteristics and vegetation cover, δ m : measurement error, ε: unmodeled error, The vector between the phase centers of the antenna looking up and the antenna looking down, θ: the elevation angle of the GNSS satellite, The slope angle of the mountain reflecting surface.

9. The method according to claim 2, characterized in that Build a GNSS-R altimetry inversion model for UAVs suitable for cities, including: Obtain the training data and label data for the constructed model; Preprocessing the model training data to obtain model training input data; Inputting the model training input data into the model to be trained for forward propagation to obtain a measured height prediction value; Calculating a loss function for model training based on the measured height prediction value and the label data for model training; Based on the loss function, backpropagation is performed on the training model until the training of the model is completed, and a height measurement inversion model suitable for the city is obtained.

10. The method according to claim 9, 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 height value of the aircraft obtained by actual measurement. 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