An ultra-low airspace three-dimensional radio map generation system

Abstract

The application provides a kind of ultra-low airspace three-dimensional radio map generation system, the system includes: three-dimensional grid division and data preprocessing module, for the target area is divided into containing longitude and latitude and height coordinate voxel grid;Satellite remote sensing image feature extraction module, for generating the environment high-dimensional image features of target area;Communication base station parameter encoder, for generating the base station space propagation characteristics of target area;Two-dimensional feature migration and three-dimensional feature generation module, for generating initial three-dimensional feature representation;Three-dimensional interpolation and height correction module, for outputting corrected three-dimensional RSRP feature;Multi-modal fusion and decoding module, for fusing corrected three-dimensional RSRP feature, environment high-dimensional image feature and base station space propagation characteristics, the fusion result is input into three-dimensional decoding network, to generate the final three-dimensional radio signal power distribution map.The application generates three-dimensional radio map for urban low-altitude scene.

Description

Technical Field

[0001] This invention belongs to the field of radio signal detection technology, specifically relating to a three-dimensional radio map generation system for ultra-low altitude domains. Background Technology

[0002] In scenarios targeting urban ultra-low-altitude communication (0–100 meters altitude), most existing radio mapping methods still rely on two-dimensional (ground plane) coverage modeling, which is insufficient to meet the needs of wireless propagation modeling within this altitude range. Current two-dimensional radio maps typically assume that the receiving equipment is located at a fixed and low altitude (such as ground or near-ground altitude), focusing primarily on the signal propagation effects in the planar direction; therefore, the characteristics of signal coverage varying with altitude are ignored.

[0003] For emerging low-altitude communication applications (such as urban drone networking or air vehicle access networks), continuing to use only planar coverage models will produce significant errors because such models cannot reflect differences in propagation characteristics in the vertical direction (such as regaining line-of-sight signal above the top of buildings, or different attenuation trends with increasing height).

[0004] Although attempts have been made to predict 3D coverage using multi-layer 2D slice overlay or complex ray tracing simulations, these methods either discretize the continuous vertical space too coarsely or have extremely high computational costs, and still fail to fully characterize the combined effects of lateral and vertical propagation factors in dense urban environments.

[0005] Therefore, existing technologies lack effective solutions for generating 3D radio maps in ultra-low altitude urban scenarios, which leads to significant shortcomings when using 2D modeling methods in network planning and optimization for airborne users. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the aforementioned background technology and provide an ultra-low altitude 3D radio map generation system that can generate effective 3D radio maps for urban low-altitude scenarios.

[0007] The technical solution adopted in this invention is: a three-dimensional radio map generation system for ultra-low altitude domains, comprising:

[0008] A 3D mesh generation and data preprocessing module is used to divide the target area into a voxel mesh containing latitude, longitude, and altitude coordinates, and to map sparse measured RSRP data to the corresponding voxel cells; the voxel mesh includes at least a ground layer and one or more altitude layers along the vertical direction;

[0009] The satellite remote sensing image feature extraction module is used to convert satellite remote sensing images of the target area into high-dimensional environmental image features;

[0010] A communication base station parameter encoder is used to map the physical parameters of each communication base station in the target area into the spatial propagation characteristics of the base station.

[0011] The 2D feature transfer and cross-plane mapping module is used to extract radio signal features of the ground layer by combining environmental features and base station propagation information, and to extend the radio signal features of the ground layer to each height layer based on voxel grid to generate an initial 3D feature representation.

[0012] The 3D interpolation and height correction module is used to perform spatial interpolation completion on voxel units for which no measured RSRP data has been obtained, based on physical propagation priors and initial 3D feature representations; combined with sparse measured RSRP data, residual correction is performed on the interpolation results, and the corrected 3D RSRP features are output.

[0013] The multimodal fusion and decoding module is used to fuse the corrected 3D RSRP features, environmental high-dimensional image features, and base station spatial propagation features. The fusion result is input into the 3D decoding network to generate the final 3D radio signal power distribution map.

[0014] The beneficial effects of this invention are as follows: By defining a modular system, including core modules such as spatial grid preprocessing, feature extraction and encoding, 2D feature transfer to 3D, 3D interpolation correction, and multimodal fusion decoding, this invention enables radio map generation to form an end-to-end closed loop from input to output. This structured design ensures a complete data flow and feature fusion path, allowing the model to not only integrate multi-source data (images, base stations, measurement data) but also incorporate prior constraints on physical propagation, significantly improving the accuracy, generalization, and practical feasibility of 3D radio maps.

[0015] Furthermore, this invention defines the ground layer height as 0 and distributes it at equal intervals with multiple height layers, marking unmeasured voxels as "missing," thus clearly delineating the spatial index and the effective area of ​​measurement points. Through a standardized and unified grid division method, a precise spatial coordinate system is provided for subsequent interpolation completion and height correction; the explicit marking of unmeasured units facilitates efficient and targeted interpolation processing.

[0016] Furthermore, this invention uses a visual Transformer structure to perform image patch segmentation and self-attention computation on remote sensing images, capturing information on land cover categories, building heights, shadows, and terrain. The visual Transformer has a stronger ability to deeply model the semantics and spatial relationships of images, significantly improving the accuracy of image features in representing complex urban environmental information, and making subsequent feature fusion and signal prediction more accurate.

[0017] Furthermore, this invention maps base station coordinates to location embedding, and fuses parameters such as carrier frequency and transmit power into a radiation feature vector through a multilayer perceptron (MLP). The two are then combined into a spatial propagation feature. This effectively integrates base station physical parameters with spatial location information to form a spatially continuous base station propagation feature field, enabling the network to directly perceive and reasonably model base station coverage characteristics, thereby improving the accuracy of signal strength prediction.

[0018] Furthermore, this invention extracts multi-scale ground features by calling a two-dimensional radio map model, then vertically expands the two-dimensional convolutional kernel and applies multi-head cross-attention to capture the correlation of signal features at different altitudes. This effectively transfers high-quality features (land cover categories, building edges, and attenuation trends) from the trained two-dimensional model, and captures continuous feature changes in the vertical direction through cross-height layer mapping and cross-attention fusion, thereby improving the quality and continuity of the initial three-dimensional features.

[0019] Furthermore, this invention generates a theoretical trend field based on an empirical logarithmic model, performs spatial interpolation on unmeasured voxels, and then uses a trainable model to perform residual correction using sparse measurement point data. The physical trend field guides the spatial interpolation, ensuring the physical rationality of the overall propagation trend, while the data-driven residual correction accurately compensates for local errors. The synergy of these two methods significantly improves the accuracy and reliability of three-dimensional signal intensity prediction.

[0020] Furthermore, this invention utilizes a cross-attention mechanism to fuse RSRP features with environmental and base station features, embeds spatial location information through 3D location encoding, and finally combines skip connections with hierarchical decoding to recover high-resolution features. Cross-attention effectively fuses multi-source feature information, 3D location encoding enables the model to have significant spatial localization capabilities, while skip connections and hierarchical decoding restore fine ground details, comprehensively improving the spatial consistency and local detail quality of the final map.

[0021] Furthermore, this invention employs multiple loss functions (supervised regression, empirical constraints, adversarial loss, structural similarity, hierarchical smoothing, and physical prior) to optimize the network end-to-end.

[0022] The multi-dimensional loss combination optimizes different aspects of each module, ensuring not only overall prediction accuracy but also physical consistency, high-frequency detail performance, smooth transition between vertical layers, and structural accuracy, thereby improving the overall network output quality.

[0023] Furthermore, this invention utilizes an empirical logarithmic model to generate a theoretical RSRP statistical distribution, constraining the difference between base station characteristics and the theoretical statistical distribution. By imposing constraints on spatial propagation characteristics through a physical empirical model, the base station propagation characteristics are ensured to remain consistent with physical laws, avoiding unreasonable biases generated by data-driven models and enhancing physical interpretability and stability.

[0024] Furthermore, in generative adversarial training, this invention uses physically-assisted maps as real samples to enhance the generative network's ability to capture real radio propagation phenomena (line-of-sight, waveguides, and shadow effects). The adversarial mechanism enables the model to generate more realistic and detailed high-frequency signal details and shadow effects, significantly improving the map's visual realism and prediction robustness.

[0025] Furthermore, this invention calculates the mean squared error (MSE) of features between adjacent height layers and constrains network training to ensure a smooth transition in predictions between each height layer. This effectively avoids drastic, non-physical fluctuations in predictions between adjacent height layers, ensuring that the predicted signal exhibits a smooth and continuous propagation trend in the vertical direction.

[0026] Furthermore, this invention utilizes a free-space path loss model to constrain the difference between the predicted maximum height layer and the theoretical physical value. This prevents physically unreasonable results from deep networks in extreme height predictions, ensuring that the overall 3D prediction conforms to the laws of physical propagation and improving physical consistency.

[0027] Furthermore, this invention calculates the structural similarity (SSIM) between the 3D predicted shadow gradient map and the satellite image building outline gradient map, and uses this as a loss function constraint. This accurately aligns the predicted radio signal shadows with the building outlines, effectively improving the spatial accuracy of the shadow effect and ensuring that the prediction results are highly consistent with the real building environment.

[0028] This invention works synergistically across multiple levels, including module partitioning, data processing mechanisms, physical constraints, and feature fusion: it reduces the need for 3D data through knowledge transfer and accelerates training using 2D priors; it combines physical models and data-driven approaches to ensure the rationality of 3D signal attenuation; and it employs a multimodal attention mechanism to effectively fuse environmental, base station, and measurement data, thereby improving the accuracy of low-altitude scene modeling. This invention significantly improves accuracy, generalization, physical consistency, detail reproduction, and spatial continuity, thus comprehensively enhancing the practical application value of ultra-low-altitude 3D radio map generation systems. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0030] Figure 2 This is a schematic diagram of the training process of the present invention;

[0031] Figure 3 For urban ultra-low-altitude wireless signal propagation scenarios;

[0032] Figure 4 This is a schematic diagram of a typical low-altitude propagation scenario in urban areas. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments to facilitate a clear understanding of the present invention, but these descriptions do not constitute a limitation on the present invention.

[0034] like Figure 1 As shown, this invention provides a three-dimensional radio map generation system for ultra-low altitude domains, characterized in that the system outputs a three-dimensional radio map representing the planar distribution information and vertical height information of radio signal power within a target area. The system includes:

[0035] A 3D mesh generation and data preprocessing module is used to divide the target area into a voxel mesh containing latitude, longitude, and altitude coordinates, and to map sparse measured RSRP data to the corresponding voxel cells; the voxel mesh includes at least a ground layer and one or more altitude layers along the vertical direction;

[0036] The satellite remote sensing image feature extraction module is used to convert satellite remote sensing images of the target area into high-dimensional environmental image features;

[0037] A communication base station parameter encoder is used to map the physical parameters of each communication base station in the target area into the spatial propagation characteristics of the base station.

[0038] The 2D feature transfer and cross-plane mapping module is used to extract radio signal features of the ground layer by combining environmental features and base station propagation information, and to extend the radio signal features of the ground layer to each height layer based on voxel grid to generate an initial 3D feature representation.

[0039] The 3D interpolation and height correction module is used to perform spatial interpolation completion on voxel units for which no measured RSRP data has been obtained, based on physical propagation priors and initial 3D feature representations; combined with sparse measured RSRP data, residual correction is performed on the interpolation results, and the corrected 3D RSRP features are output.

[0040] The multimodal fusion and decoding module is used to fuse the corrected 3D RSRP features, environmental high-dimensional image features, and base station spatial propagation features. The fusion result is input into the 3D decoding network to generate the final 3D radio signal power distribution map.

[0041] Specifically, the 3D mesh generation and data preprocessing module defines the ground layer height as 0, and the ground layer and multiple height layers are distributed at equal intervals; and voxel units without radio reference signal power measurement values ​​are marked as "missing".

[0042] Preferably, the 3D mesh generation and data preprocessing module can complete the 3D mesh generation and data preprocessing in the following manner:

[0043] 1) Coordinate system and resolution settings

[0044] Select a planar projection coordinate system (such as UTM or Gauss-Kruger projection) and convert all latitude and longitude coordinates within the target area into planar X / Y coordinates;

[0045] Determine the horizontal resolution Δx and Δy (e.g., 5m × 5m) and the vertical resolution Δh (e.g., 10m) based on the project requirements.

[0046] 2) Voxel mesh construction

[0047] The transformed plane interval [X] min ,X max ]、[Y min ,Y max The area is divided into H×W planar grid cells according to Δx and Δy, where the planar grid cells are the two-dimensional grids corresponding to the "ground layer" (height h=0);

[0048] In the vertical direction, starting from the ground layer (h=0m), N–1 height layers (h=Δh,2Δh,…,(N–1)Δh) are generated sequentially at intervals of Δh, forming a total of N×H×W three-dimensional voxels;

[0049] 3) Sparse measurement point mapping

[0050] For each RSRP measurement point collected in the field, its latitude and longitude are first converted to plane X / Y, and then the index (i,j,k) falling into the grid is calculated using subscripts:

[0051]

[0052] The measured RSRP value is written into the corresponding voxel cell. If multiple measurement points fall into the same cell, their average or maximum value can be taken.

[0053] 4) Missing markers

[0054] During initialization, the RSRP values ​​of all N×H×W voxel units are set to the "missing" flag;

[0055] Only the voxel units corresponding to the sparse measurement points obtained by the above steps are replaced with the actual RSRP measurement values;

[0056] 5) Data Structures and Storage

[0057] Store the grid and its attributes as a three-dimensional array or tensor; at the same time, save the corresponding three one-dimensional coordinate vectors X, Y, and h for spatial indexing and visualization in subsequent modules;

[0058] 6) Boundary and validity checks

[0059] Perform boundary pruning on the mapping index (i,j,k) to ensure that all measurement points fall within [0,H–1]×[0,W–1]×[0,N–1].

[0060] Report or remove any measurement points that exceed the boundaries or whose height exceeds the range;

[0061] Through the above steps, the system can generate a preliminary three-dimensional RSRP mesh data of N×H×W, where the "missing" marker explicitly indicates voxel units without measurement values, providing input basis for subsequent spatial interpolation and height correction modules.

[0062] Specifically, the satellite remote sensing image feature extraction module adopts a multi-layer visual Transformer structure. By dividing the input satellite remote sensing image into a series of image patches and performing a self-attention mechanism on each image patch, it obtains spatial-semantic fusion features that simultaneously reflect land cover categories, building heights and shadow distributions, and ground elevation as high-dimensional environmental image features.

[0063] Preferably, the satellite remote sensing image feature extraction module is mainly used to extract multi-scale spatial features from high-resolution satellite images. The input includes high-resolution satellite remote sensing images (such as aerial remote sensing images) covering the target area. The output is a multi-scale feature representation tensor containing information such as the distribution of land cover categories, building heights, and / or ground elevation distribution. The core processing logic utilizes deep neural networks to encode the image features, typically using a convolutional neural network (CNN) or visual Transformer architecture. In specific implementations, through multi-layer convolution and pooling (downsampling), or a self-attention-based feature extraction mechanism, the original image is progressively mapped into feature maps of different scales and semantic levels: containing both fine-scale land cover edge contour features and global environmental context within a larger receptive field.

[0064] The extracted features can characterize key information such as the category of surface objects (e.g., buildings, roads, vegetation), shape boundaries, and corresponding building height distribution and ground elevation distribution. Key parameters include convolutional kernel size, number of network layers, feature channel dimension, and input image resolution. These parameters determine the accuracy and multi-scale expressiveness of feature extraction; for example, the convolutional kernel and hierarchical design affect the ability to capture small targets and large structures. The multi-scale feature tensors extracted by the satellite remote sensing image feature extraction module will serve as one of the inputs for subsequent multimodal fusion, providing rich environmental perception information for radio map generation.

[0065] Specifically, the communication base station parameter encoder is used for:

[0066] The geographic coordinates of the base station are mapped to a grid-aligned location and embedded.

[0067] The spatial radiation feature vector is output by combining the location, carrier frequency, antenna transmit power, vertical downtilt angle and / or base station height.

[0068] The spatial propagation features of the base station are generated by fusing the location embedding and spatial radiation feature vectors.

[0069] Preferably, such as Figure 2 As shown, the communication base station encoder consists of a base station physical feature input module and a base station parameter decoder. Its input also includes random noise to enhance the diversity and robustness of the system during training and inference.

[0070] The base station physical feature input module is used to input the physical parameters of the base station (such as location, carrier frequency, antenna transmit power, vertical downtilt angle, and / or base station height) into the system, map or encode them into a latent feature vector, and input this vector into the multimodal fusion module. This feature representation is typically a "high-dimensional vector" required for subsequent network fusion or decoding, designed to capture key information about the base station's impact on wireless signal coverage. The input is a set of parameters from one or more communication base stations, including key parameters such as the geographic latitude and longitude coordinates, antenna installation height, operating frequency band, transmit power, and downtilt angle (elevation angle) for each base station. To facilitate subsequent network utilization, these raw parameters are usually normalized or preprocessed. The output is a fixed-length base station parameter embedding vector, which represents the comprehensive features of the base station with continuous numerical values. The core processing logic uses a multilayer perceptron (MLP) to perform nonlinear transformation encoding on the input base station parameters.

[0071] Specifically, the base station physical feature input module extracts the correlation between different parameters step by step through several fully connected layers: the first layer maps each parameter to a high-dimensional space to provide sufficient expressive power; subsequent hidden layers introduce nonlinearity through activation functions (such as ReLU), enabling the combined feature vector to characterize complex relationships such as the impact of base station location and frequency band on coverage. The final output layer generates a fixed-dimensional vector as the encoding result. Key parameters include the number of layers in the MLP, the number of neurons in each layer (hidden layer dimension), the type of activation function, and the dimension of the output embedding vector. These design parameters determine the richness and refinement of the base station physical feature input module's representation of base station features. For example, the output vector dimension needs to be large enough to cover the differences in the impact of latitude and longitude location on the coverage area, but not too large to avoid introducing redundancy. The embedding vector generated by the base station physical feature input module is provided to the multimodal fusion module, working in conjunction with environmental features and measurement signals: during the fusion process, features from different spatial locations are combined with this vector, allowing the model to perceive the attributes of a specific base station (such as transmit power or frequency band differences) and adaptively adjust the interpretation of environmental and signal features. This enables the effective fusion of base station information with other modal data, ensuring that the generated radio map is specific to the current base station configuration.

[0072] The base station parameter decoder is based on a diffusion model or other generative framework. It uses random noise and the latent vector z output by the base station physical feature input module to perform a "reverse denoising" or "generation" process, and finally outputs the estimated base station parameters. During training, it compares the parameters with the real base station parameters to generate a loss, thereby constraining the encoder's mapping process.

[0073] The forward diffusion process of the base station parameter decoder includes: gradually adding noise to the data (base station parameters) to move from the initial true values ​​to a Gaussian distribution (or other noise distribution); this can be regarded as a multi-level perturbation of the base station parameters in offline or training mode to simulate uncertainty or measurement error.

[0074] The reverse generation process of the base station parameter decoder includes: a diffusion-based denoising network starts with pure noise or partially noisy data, iteratively denoises, and finally gradually restores clean base station parameters. This denoising process is mutually constrained or fused with the latent vector z output by the base station physical feature input module to ensure that the restored result matches the true parameters.

[0075] At the base station parameter decoder, random noise is typically introduced in the following ways:

[0076] Initial noise sampling: During back diffusion, a noise tensor is first sampled from a standard normal distribution, and then base station parameters are generated in reverse through a multi-step denoising network;

[0077] Fusion encoder features: Several intermediate steps can inject the latent vector z output by the base station physical feature input module into the decoding network, so that the decoder can simultaneously consider the uncertainty of "existing physical prior information z" and "random noise" to generate the final approximate value of base station parameters.

[0078] Reconstruction loss: By comparing the decoder output with the real base station parameters and minimizing the reconstruction error, the base station physical feature input module and the decoder are forced to learn the correct denoising mapping during training and form an accurate restoration of the base station parameters.

[0079] During the training process of the communication base station parameter encoder, after the base station physical feature input module maps the base station physical parameters to the latent vector z, the base station parameter decoder generates a reconstructed base station parameter p from z and random noise through a diffusion model for inverse denoising. If the difference between p and the true parameters is too large, a high reconstruction error (such as mean square error) will occur, which propagates back to the base station physical feature input module, forcing it to learn a more physically meaningful and reversible representation. For the base station parameters of each training sample, noise can be added forward (simulating measurement errors, diverse scenarios), and then the base station parameter decoder can be used to gradually denoise in reverse. If the denoised result is still close to the original parameters, it indicates that the model has successfully captured the statistical distribution of the base station parameters and tolerates a certain degree of random perturbation. Base station parameters are often limited by engineering constraints in reality (such as the transmit power being within a certain range, and the downtilt angle not being too large). The diffusion model can reflect these priors in the latent space or denoising process. If the decoded result exceeds the reasonable domain, a loss penalty will be used to guide the base station physical feature input module-base station parameter decoder to adjust to a range that conforms to physical laws.

[0080] Compared to purely deterministic mappings, diffusion models allow for the preservation of some randomness during the inverse denoising stage, enabling slightly different samples to be generated for the same base station input parameters during training. This simulates minor measurement / input errors in base station deployment, improving model generalization. When training on single-scene or local data, without noise injection and generative mechanisms, the encoder may overfit a specific input-output mapping; with the bidirectional constraint of diffusion denoising, the network is forced to learn the base station parameter distribution more broadly, thus handling the inevitable minor variations in the real environment during inference.

[0081] In actual training, the reconstruction loss generated by the base station parameter decoder often appears together with other loss terms (such as empirical prior loss, adversarial loss, and measured RSRP regression loss).

[0082] During the use of a base station encoder, the base station parameter decoder is used to generate or recover base station parameters. To verify the encoder's fidelity in reproducing base station parameters, or to complete missing / incomplete base station parameters in certain scenarios, the decoder can perform inverse denoising starting from the noise plus z, generating a list of feasible base station parameter candidates. Sometimes, engineering projects require simulating more "virtual base stations" or "base station configurations" to observe coverage changes. The decoder can sample different noise levels, resolving diverse base station parameter instances, which is helpful for large-scale simulation testing.

[0083] The base station parameter decoder is also used for uncertainty quantification: given the same latent vector z, the decoder can sample and denoise noise multiple times during inference to obtain slightly different p, thereby estimating the uncertainty range of the base station parameters. If deviations are found in certain base station parameters (such as antenna height) during actual deployment, the decoder can be updated by retraining or fine-tuning so that the new latent vector reflects this adjustment, thus maintaining the system's accurate representation of the physical environment.

[0084] The base station parameter decoder is primarily used in conjunction with radio map generation: after confirming that the latent vectors of the base station parameters are consistent with the actual parameters, this "physically consistent" z-vector is fused with environmental features and input into the radio map decoder, ultimately generating a more reliable coverage map. When operators change base station configurations (frequency band, power, downtilt angle, etc.), the decoder can also perform denoising under the new parameters, ensuring that the new latent vectors are fused with the environment to output updated coverage predictions, facilitating online or offline planning and evaluation.

[0085] Specifically, the two-dimensional feature transfer and cross-plane mapping module is used for:

[0086] The pre-trained 2D radio map generation model is invoked to extract multi-scale convolutional features of the ground layer, including the distribution of land cover categories, building edges, and signal attenuation trends, by taking high-dimensional environmental image features and spatial propagation features of base stations as input.

[0087] The two-dimensional convolutional kernel parameters used to extract multi-scale features in the two-dimensional radio map generation model are copied and expanded into three-dimensional convolutional kernels in the vertical direction. The three-dimensional convolutional kernels are applied in parallel to each multi-scale feature map of the ground layer at the corresponding height layer to generate preliminary three-dimensional feature slices for each height layer.

[0088] The initial 3D feature slices from each height layer are input into a multi-head cross-attention network, which fuses feature representations from different height layers to capture vertical correlations and output the initial 3D feature representation.

[0089] Preferably, the two-dimensional feature transfer and cross-plane mapping module can implement two-dimensional feature transfer and cross-plane mapping in the following manner:

[0090] 1) Multi-scale feature extraction from two-dimensional models

[0091] High-dimensional environmental image features Spatial propagation characteristics of base stations By splicing along the channel dimension, we obtain .

[0092] Using a pre-trained two-dimensional radio map generation network (e.g., based on ResNet or VisionTransformer, such as the model described in prior patent application CN202411613716.9, "A Method for Generating Radio Maps and Object Distribution Maps, Switching Models and Training Methods Thereof"), input F at the ground layer (height=0). in ;

[0093] In its final convolutional stage or Transformer layer, multi-scale convolutional features are extracted in parallel:

[0094] Small receptive field: 3×3 kernel, number of channels;

[0095] Mid-receptive field: 5×5 kernel, number of channels;

[0096] Large receptive field: 7×7 kernel, number of channels;

[0097] Output three sets of two-dimensional feature maps They capture information such as the distribution of land cover types, building edges, and decay trends.

[0098] 2) Vertical convolution kernel expansion

[0099] For each receptive field, the parameters of the two-dimensional convolution kernel are... Copy to 3D convolution kernel Make In the initial state, the feature extraction capability in the horizontal direction remains unchanged, and the receptive field is only formed in the depth (height) dimension.

[0100] For each set of features, apply the corresponding 3D convolution in parallel along the height dimension.

[0101] Obtain a preliminary set of 3D slices for each height layer. .

[0102] 3) Layered cross-attention fusion

[0103] Multi-scale slices from each depth layer are stitched together according to channels to form .

[0104] Generate a learnable height code p for each voxel location (i,j,h). h ∈R D and with X hAdd (i,j) together.

[0105] For all layers of features {X h +p h Simultaneously serving as Query, Key, and Value, it utilizes N attention heads for parallel computation:

[0106]

[0107] Where Q=X h +p h K=V=[X1+p1;… ;X Z +p Z ]

[0108] Output the fused features of each layer

[0109] All Concatenate the data according to the depth layer dimension to generate an initial 3D feature representation. .

[0110] Specifically, the three-dimensional interpolation and height correction module is used for:

[0111] Based on the empirical logarithmic fitting model, a theoretical RSRP trend field of the same size as the initial three-dimensional features is generated;

[0112] By combining the trend field, initial three-dimensional features, and sparse measured RSRP data, missing voxels are filled using spatial interpolation methods.

[0113] By combining sparse measured RSRP data, a trainable model is used to perform residual correction on the interpolation results, and the corrected three-dimensional RSRP features are output.

[0114] Preferably, the three-dimensional interpolation and height correction module can implement three-dimensional interpolation and height correction in the following manner:

[0115] 1) Generation of physical trend fields

[0116] The initial three-dimensional features are averaged by channel at the ground layer (h=0) to obtain P0(x,y).

[0117] Applying the modified Okumura-Hata model to each altitude level, the theoretical signal attenuation value is calculated:

[0118]

[0119] Where h is the current height of the altitude layer, and h reference is the ground reference height.

[0120] This leads to the theoretical trend field. The tensor mentioned above has the same dimensions as the initial three-dimensional feature and can be directly aligned voxel by voxel.

[0121] 2) Residual field construction and 3D kriging interpolation

[0122] For each measured data (x) i ,y i ,h i ,P meas,i Calculate the residuals at the measuring points: .

[0123] Let the horizontal distance d and the height difference Δh be defined, and the anisotropic variability function be defined as follows:

[0124] Where c0, c1, α, β are obtained by fitting the experimental variogram.

[0125] Calculate the weight λ for the estimated voxel (x0, y0, h0) i

[0126]

[0127] Where, γ ij =γ(d ij ,Δh ij ), γ i0 =γ(d i0 ,Δh i0 );γ ij γ represents the semivariance between the residuals of the i-th and j-th measurement points; i0 denoted as , representing the semivariance between the residual at the i-th measurement point and the point to be estimated; μ represents the Lagrange multiplier.

[0128] The residual interpolation is superimposed with the theoretical trend field to obtain the Kriging interpolation tensor. :

[0129] ,

[0130]

[0131] Kriging Results An estimate is given for all cells.

[0132] 3) Data-driven residual correction

[0133] Kriging interpolation tensor and the sparse tensor of the residual at the measuring point ;

[0134] As a two-channel input to a trainable network (lightweight CNN, such as U-Net), the output is a residual correction tensor. .

[0135] The final output is the corrected 3D RSRP feature. .

[0136] In this embodiment, during the residual correction stage, the Kriging interpolation tensor and the 'measurement point residual mask' are concatenated along the channel dimension and fed into a lightweight U-Net. This U-Net acts as an independent sub-network, using only the MSE (or MAE) at the measurement points as the target during training, and does not participate in the joint backpropagation of the backbone network. After training and freezing the weights, only one additional forward pass is needed during the inference stage to generate a voxel-level residual map, which is then added to the interpolation result to obtain the final correction value. Due to the shallow model depth and small number of channels, the impact on the overall computational cost is controllable; simultaneously, it can learn nonlinear residual distributions, improving the local accuracy of unmeasured regions, and the output can still be used with subsequent fusion or adversarial branches.

[0137] Specifically, the multimodal fusion and decoding module further includes: a multimodal fusion module and a decoder.

[0138] The multimodal fusion module includes:

[0139] The cross-attention fusion layer uses the corrected 3D RSRP features as the query vector, and the high-dimensional environmental image features and the base station spatial propagation features as the key / value vectors. Multiple attention heads are processed in parallel to generate fused features.

[0140] The 3D position encoder maps the 3D coordinates of each voxel to a spatial embedding of the same dimension as the fused feature, and adds it to the fused feature element by element, enabling attention computation to have absolute spatial awareness.

[0141] The decoder takes the ground layer features output by the 2D feature transfer module and introduces them into the corresponding upsampling stage of the decoder through skip connections. It then upsamples and fuses the features layer by layer according to the planar resolution and height interlayer spacing of the voxel grid, and outputs the final 3D radio signal intensity distribution map.

[0142] Preferably, the cross-attention fusion layer achieves the above function using the following steps:

[0143] The corrected 3D RSRP feature tensor (size is height × width × depth × number of channels) is first linearly transformed to map it to a fixed feature dimension (e.g., 256 dimensions), which serves as the query for multi-head attention.

[0144] The high-dimensional environmental image features and the spatial propagation features of the base station are mapped to the same feature dimension through their respective linear mappings, and then concatenated or added by channel to serve as the key and value for attention.

[0145] This module internally sets up several (e.g., 8) parallel attention heads, each of which uses an independent linear mapping to project the Query, Key, and Value to a lower-dimensional subspace (e.g., 32-dimensional).

[0146] Within each head, the dot product of Query and Key in the channel dimension is first calculated to measure the correlation of different modal features at each voxel position;

[0147] The dot product result is normalized (Softmax) to obtain the attention weights, and then a weighted sum is performed with the corresponding Value to obtain the output of each head;

[0148] The outputs of all heads are concatenated back to the original feature dimensions, and then integrated into a fused feature through a linear layer;

[0149] A residual connection is retained before and after the cross-attention operation to ensure that the main branch information is not lost due to attention shift; after each attention or linear transformation, a layer normalization layer (LayerNorm) is immediately followed to stabilize training;

[0150] The final output is a fused feature tensor of the same size as the input. In terms of spatial location, it contains both the vertically corrected signal information and the propagation priors of the environment and the base station.

[0151] Preferably, the three-dimensional position encoder achieves the above functions by means of the following steps:

[0152] For each voxel unit, its horizontal index (i, j) and height index h in the mesh are extracted and mapped to the range [0,1] by normalization (divided by the maximum width / height minus one);

[0153] The normalized 3D coordinate vector is input into a small multilayer perceptron (MLP), which contains two layers of linear transformation and activation functions, and outputs a spatial embedding vector with the same dimension as the fused features (e.g., 256 dimensions).

[0154] The fusion features obtained by embedding the space and the cross-attention fusion layer are added element-wise in the channel dimension to ensure that the final representation of each voxel carries both content and location information.

[0155] In subsequent attention or convolution operations, the network can use absolute spatial location to distinguish the different propagation characteristics of the same behavior at different floors or plane locations. For example, the physical meaning of the same shadow pattern will be different on high floors and ground floors.

[0156] Preferably, the decoder consists of several (e.g., 4 to 5) 3D transposed convolution modules arranged in sequence. Each layer upsamples the height, row, and column directions simultaneously or stepwise, gradually reconstructing the fused features from the coarse low resolution to the original horizontal resolution and height interlayer spacing of the target voxel mesh.

[0157] Each transposed convolutional unit includes:

[0158] A 3×3×3 or 4×4×4 3D transposed convolutional layer with a stride of 2 typically expands the size of the three-dimensional voxels by about 2 times in each of the three dimensions.

[0159] This is followed by a batch normalization (BatchNorm) and nonlinear activation (ReLU);

[0160] If necessary, a 3×3×3 conventional 3D convolutional layer can be added to refine the features.

[0161] For example, the first layer magnifies the coarse fused feature voxels by a factor of 2 in each of the three directions, and the next layer continues to magnify them until they are restored to the same size as the original voxel mesh in terms of horizontal resolution and height interlayer spacing.

[0162] After the first or second level of upsampling is completed, when the planar resolution of the decoder is close to the original resolution of the ground layer, high-resolution ground layer features output by the two-dimensional feature transfer module are introduced.

[0163] During each upsampling stage of the decoder, when the resolution of the current voxel in the height direction first approaches the planar resolution corresponding to the "ground layer," the aforementioned 2D feature map is aligned with the decoded features as follows: the 2D feature map is upsampled / downsampled or padded as necessary to make its row and column count consistent with the size of the current 3D feature in both planar dimensions; this 2D feature map is copied or broadcast in the height direction as a 3D volume with a small depth, and then concatenated or added to the current decoded feature tensor in the channel dimension. This skip connection allows the decoder to both utilize the 3D structural information captured by the deep fusion features and quickly supplement lost planar details such as ground edges and road textures.

[0164] After each stitch, a lightweight 1×1×1 3D convolution or channel-weighted operation can be used to merge the channel dimensions of the stitched data back to a uniform size, and an additional 3D convolution with a small receptive field can be used to refine the features and eliminate the channel redundancy caused by "stereotyping".

[0165] After the final upsampling stage, the decoder compresses the multi-channel features into a single-channel tensor using a 1×1×1 3D convolution, outputting an RSRP prediction value for each voxel. A linear activation layer can be added as needed to map the network's original output to the actual power range.

[0166] Through the above process, the decoder not only ensures continuous upsampling of features at all levels in three-dimensional space, but also cleverly utilizes two-dimensional ground features to enhance planar details and restore a high-precision, hole-free three-dimensional radio signal strength distribution map.

[0167] Specifically, the present invention also includes a three-dimensional propagation mechanism modeling module. By introducing explicit modeling of the components of wireless signal propagation into the model, the physical credibility of the final generated radio map and the adversarial training effect are improved. This module generates corresponding propagation component maps for three main mechanisms: line-of-sight propagation, waveguide propagation, and shadow fading, so as to form the auxiliary label information required for adversarial loss and structural loss during training. Specifically, the three-dimensional propagation mechanism modeling module can add three parallel branches after the decoder or feature fusion, each branch focusing on characterizing the signal distribution changes caused by a certain propagation phenomenon.

[0168] like Figure 4 As shown, the line-of-sight propagation diagram can simulate the direct ray path assumption between the base station and a spatial point. It can usually be approximated by a small convolution kernel and a shallow network, and the free space path loss will decay at a relatively stable rate once the distance and height are determined.

[0169] In this embodiment, the three-dimensional propagation mechanism modeling module establishes a path loss model for line-of-sight wireless transmission users affected by the superposition of primary reflected waves and direct waves from the ground and walls at close range, based on the multipath channel formed by the superposition mechanism of reflected waves from the ground and walls and direct waves:

[0170] Based on the principle of three-dimensional mirror reflection, a system including vertical height differences is established. The propagation path model. Let the spatial coordinates of the base station and the UAV be... and The ground reflection point and the wall reflection point satisfy the following:

[0171]

[0172] in Angle of incidence at ground level The azimuth angle reflected from the wall surface. The width of the street.

[0173] dx=|x u -xBS∣、dy=∣y u -y BS|: The projected distance between the base station and the drone on the horizontal plane along the x and y directions;

[0174] Δh=|h UAV -h BS |: Vertical height difference between the base station and the drone;

[0175] d Ref1 : Total length of the ground reflection path;

[0176] d Ref2 : Total length of the reflection path on the wall.

[0177] Compared to the two-dimensional model that only considers horizontal reflection, this embodiment introduces three-dimensional characteristic enhancement by considering the impact of the receiver height on path loss. λ represents the signal wavelength, and the subscript i=1 corresponds to ground reflection, i=2 corresponds to wall reflection.

[0178] ① Vertical phase compensation factor: Add a phase difference correction term for the reflection path. The impact of quantization height difference on multipath interference.

[0179] ② Spatial correlation modeling: Define the three-dimensional spatial coherence radius as This characterizes the impact of UAV altitude changes on the correlation of multipath signals.

[0180] The straight-line distance between the base station and the drone on a plane.

[0181] Therefore, the path loss formula for the line-of-sight multipath coherent stacking model can be described as:

[0182]

[0183] Among them, the three-dimensional interference enhancement factor is:

[0184]

[0185] This indicates the line-of-sight (LOS) distance between the base station and the drone;

[0186] f represents the signal frequency (unit: GHz);

[0187] Γ i This represents the field reflection coefficient (complex amplitude) of the i-th reflection path.

[0188] like Figure 4 As shown, waveguide propagation maps can be used to characterize the along-path enhancement effect in narrow environments or complex terrains. They may capture the long-distance extension of signals in the horizontal dimension and multiple refractions and reflections through a large receptive field or long sequence attention mechanism.

[0189] The 3D propagation mechanism modeling module establishes a series expansion model of 3D waveguide propagation for users at the end of a street, considering the multiple reflections and superpositions caused by buildings on both sides of the street. It assumes a propagation distance of d along the street axis, a maximum effective number of reflections N, and an angle between the waveguide propagation direction and the street axis. Considering the combined field strength along the ground reflection path:

[0190]

[0191] Wherein, E0: reference incident field strength amplitude (direct wave amplitude without reflection).

[0192] Γ: Amplitude of the field reflection coefficient of a single reflection (loss factor when reflected from the ground or wall).

[0193] k: Reflection level index, from 0 (direct wave) to NNN (maximum number of reflections).

[0194] N: Maximum number of effective reflections, limited by street length and angle of incidence (see formula below).

[0195] d: The horizontal projection distance of the direct or first reflected wave along the street axis.

[0196] Δs: The increase in propagation path between adjacent reflections (see the formula below).

[0197] λ: wavelength.

[0198] Assuming the wall and ground reflections are symmetrical, the increment of adjacent reflection path lengths... Maximum effective number of reflections .

[0199] Where w: the "street width" or the distance between reflective surfaces between buildings on both sides of the street.

[0200] φ: Angle between the waveguide propagation direction and the street axis (angle of incidence)

[0201] when When considering the finite number of reflections and phase randomness of actual streets, the simplified two-dimensional waveguide gain is obtained:

[0202]

[0203] Compared to a two-dimensional model that fixes the number of reflections, a three-dimensional model introduces a height ratio correction factor. The effect of quantization terminal height on waveguide mode modulation. The corrected effective number of reflections in three-dimensional space is:

[0204]

[0205] The final expression for the three-dimensional waveguide gain is:

[0206]

[0207] The vertical dimension gain correction term is added:

[0208] .

[0209] like Figure 4 As shown, the shadow fading map can target slow fading caused by buildings and terrain undulations, tending to output areas of wide-area signal attenuation. Lower-resolution convolution or random perturbations can be used to simulate statistical characteristics, reflecting the segmented fading patterns caused by building occlusion.

[0210] The 3D propagation mechanism modeling module addresses users affected by the shadowing effect caused by building occlusion. It establishes a low-altitude building occlusion model based on multiple diffraction and considering shadow fading. Based on the theory of uniform diffraction, the 3D diffraction field strength can be expressed as:

[0211]

[0212] Where E0: the magnitude of the incident field strength at the diffraction point along the direct path (field strength without diffraction).

[0213] D(θ i ,φ i ): Diffraction coefficient, which depends on the incident angle and the azimuth angle along the edge direction, and is used to simulate the diffraction gain and phase of the incident wave by the edge.

[0214] ρ: The distance from the launch point to the edge of the diffraction.

[0215] s: The distance from the edge of the diffraction to the receiving point.

[0216] k: wave number, k = 2π / λ.

[0217] λ: Signal wavelength.

[0218] θ i The angle of incidence (vertical profile) between the incident wave and the plane containing the diffracted edge.

[0219] φ i : The azimuth angle of the incident wave in the plane of diffraction edge.

[0220] Diffraction coefficient for:

[0221]

[0222] : Smooth transition function

[0223] Diffraction parameters;

[0224] a+, a-: Geometric distance or angular offset from the incident / outgoing paths on either side of the diffraction edge.

[0225] θ d : Diffraction exit angle.

[0226] By using the Deygout method to calculate step-by-step for the M diffraction edges, we can obtain:

[0227]

[0228] Where M: the total number of effective diffraction edges (reflection levels).

[0229] d m : The total length of the m-th diffraction / reflection path.

[0230] h m The height of the building or obstacle corresponding to the m-th diffraction edge.

[0231] For 3D correction items:

[0232]

[0233] Among them, h building The height of a building that causes shadows or diffraction.

[0234] d: Street width or spacing between reflective surfaces (horizontal distance).

[0235] θ v The elevation angle (or antenna downtilt angle) of the incident wave on the vertical profile is used to quantify the effect of altitude on the waveguide mode.

[0236] Finally, we model shadow fading based on a probabilistic model:

[0237]

[0238] in, Based on the baseline shadow standard deviation, a log-normal distribution is statistically fitted using measured data to reflect the impact of building density and height dispersion on the signal. A vertical correction factor is then added. Quantifying the additional effects of vertical building occlusion in three-dimensional space, when When the signal is at a low elevation angle, it propagates approximately horizontally, and the shadow fading characteristics are close to a two-dimensional model; when When the signal is at a high elevation angle and perpendicularly passes through a cluster of buildings, shadow fluctuations will be amplified. Finally, a synthesis model for the total path loss is presented:

[0239]

[0240] Where d: Transmitter-receiver direct range (km).

[0241] L diff,m : Loss term for the m-th diffraction path.

[0242] χ σ : Follows the zero mean and zero variance σ 3D 2 Gaussian random variables are used to simulate shadow decay.

[0243] During forward propagation, this module, based on the aforementioned modeling of the 3D propagation mechanism modeling module, generates corresponding 3D component maps for each propagation branch, each aligned with the real geographic coordinates. In adversarial training and structure supervision, the line-of-sight, waveguide, and shadow component maps can be integrated into a joint or integrated auxiliary propagation label. This involves combining the 3D maps corresponding to the three mechanisms in a weighted or spliced ​​manner, serving as the "real propagation component combination" as input to the discriminator (or discrimination criterion).

[0244] In practical implementation, the signal distribution in a real-world scenario can be approximated by pre-simulation or live network measurements into corresponding line-of-sight, waveguide, and shadow components, and then fused into a higher-dimensional or channel-specific "real label." If the three components generated by the model differ significantly from this combined label in statistical characteristics or spatial distribution, the discriminator will output an adversarial error, thereby guiding the model to continuously correct the simulation results of the three propagation mechanisms during training.

[0245] In terms of adversarial loss, the integrated auxiliary label acts as a "real sample," comparing the model-generated three component combinations (i.e., "fake samples") to determine its realism. If the model-generated view distance decay does not conform to free space rate, or if the shadow area is misaligned with the real occlusion area, the discriminator can learn the discriminative features and feed them back to the generator network, allowing it to further adjust the parameters of the view distance branch, shadow branch, etc. This enhances the generalization ability to complex environments, making the final synthesized RSRP prediction map closer to the real scene in both macroscopic trends and local shadow effects.

[0246] Regarding structural losses, this module defines corresponding physical structural constraints for each component map. For line-of-sight propagation maps, the error of the free-space attenuation formula can be used as a metric; for waveguide propagation maps, piecewise attenuation or gradient smoothness constraints can be established by combining horizontal extension and refraction effects; and for shadow fading maps, morphological matching is performed using building locations or statistical occlusion regions. During model training, the structural losses of these three components, along with the fused integrated auxiliary label adversarial loss, can be added to the overall objective function.

[0247] Through the above design, the 3D propagation mechanism modeling module of this invention achieves explicit differentiation and fusion of the three mechanisms of line-of-sight, waveguide, and shadow during the training phase, and combines the three components into an auxiliary label for adversarial discrimination. This allows the overall generated model to be simultaneously constrained by the morphology of local components and guided by global adversarial realism. On the one hand, the model output is more consistent with actual physical phenomena in each propagation mechanism; on the other hand, complex effects such as building occlusion, urban canyon waveguides, and long-distance line-of-sight in 3D maps are reflected more accurately, thereby significantly improving the spatial detail and physical reliability of the prediction results.

[0248] Specifically, the training process of this system employs a supervised learning method and constructs the following combination of loss functions:

[0249] Supervised regression loss is used to penalize the deviation between the true value label and the 3D radio signal power distribution map output by the multimodal fusion and decoding modules.

[0250] Empirical loss is used to constrain the difference between the spatial propagation characteristics of the base station output by the communication base station parameter encoder and the theoretical RSRP statistical distribution;

[0251] Combat loss is used to constrain high-frequency details and shading effects in the 3D radio signal power distribution map output by the multimodal fusion and decoding module;

[0252] Structural similarity loss is used to constrain the alignment of the signal shadow region with the building outline of the target region in the 3D radio signal power distribution map output by the multimodal fusion and decoding module.

[0253] Hierarchical similarity loss is used to constrain the smooth transition of the output of the two-dimensional feature transfer and cross-plane mapping module between adjacent height layers;

[0254] The physical prior loss is used to constrain the deviation between the average predicted value of the highest-height layer output by the 2D feature transfer and cross-plane mapping module and the theoretically calculated value.

[0255] Preferably, the supervised regression loss is implemented through the following steps: using sparse ground RSRP measured data or simulation results from a high-precision simulation model as ground truth labels, the difference between the radio signal strength distribution map output by the multimodal fusion and decoding module and the ground truth labels is quantified to correct the system output. The supervised regression loss is applied to the multimodal fusion and decoding module to ensure that the RSRP map output by the model accurately fits the actual measured values ​​at the measured locations. MSE or MAE is calculated between the predicted values ​​and ground truth labels at the RSRP sampling points, and the difference is used as the supervised regression loss.

[0256] The sources of simulation labels can include:

[0257] The downlink RSRP at each voxel is calculated using a full-wave or fast ray tracing model based on an existing 3D urban model (buildings, terrain) and base station parameters.

[0258] By using semi-empirical channel simulation, and utilizing the improved COST, Okumura-Hata, and ITU macro / micro cell models, approximate values ​​are generated in batches according to different height layers.

[0259] Alternatively, based on statistical or Monte Carlo methods, dense sampling points can be deployed according to the scene, and the power range of unmeasured areas can be quickly inferred by statistically analyzing shadows and slow fading distribution.

[0260] The tag fusion strategy implemented here is as follows:

[0261] In the field measurement of RSRP, the latitude, longitude, and altitude of each measurement point were mapped to a voxel index according to the grid division rules; the power value was directly written to the corresponding voxel position. Its confidence weight was set to 1.0 (the baseline provides "hard anchor points" for the network to ensure that the predicted values ​​at these positions are absolutely close to the actual measurements).

[0262] Ray tracing / semi-empirical simulation results are used to calculate power in batches based on a 3D city model and base station parameters; the results are mapped to the mesh in the same way as the actual measurements; if there is overlap with the actual measurements, the actual measurements take precedence. The confidence weight is set to 0.3–0.5 to fill in the supervision information in the gaps between buildings, street valleys, or high-altitude layers that are missing in the actual measurements, which speeds up convergence and maintains physical consistency. The results form two sparse "label masks"—one storing the actual values ​​and the other storing the simulation values, each with a weight coefficient.

[0263] After each forward propagation, the network outputs a complete 3D power heatmap. Using two label masks, only the predicted value corresponding to the voxel is extracted from the tensor. The mean squared error (MSE) is the preferred error metric, but a small amount of MAE can be mixed in to improve robustness. Each residual is multiplied by the confidence weight of its label, and the average of all residuals is obtained to obtain a single scalar—the supervised regression loss.

[0264] The gradient of this loss only applies from the multimodal fusion layer to the end of the decoder, ensuring that the weights of other physical priors or adversarial branches are not directly affected. In the early stages of training, the weights of the simulation labels can be increased to allow the network to quickly capture global trends. As training progresses, the simulation weights can be gradually reduced or the simulation labels can be frozen, retaining only the constraints of the measured points to complete fine correction.

[0265] Based on the above training process, the error between the system network's output values ​​and the true values ​​at all measured locations is significantly reduced, forming reliable "landmarks." Simulated labels provide soft constraints for void areas, preventing predictions from drifting at high altitudes, in valleys, or hidden corners. Dense simulated supervision signals in the early stages of training significantly shorten convergence time; even with sparse measurement points, the simulated labels remain based on the real 3D scene and propagation model, avoiding non-physical anomalies in the network output. The supervised regression loss implements a dual-layer supervision strategy of "hard reality + soft simulation," ensuring both measurement point accuracy and maintaining overall spatial continuity and physical plausibility, ultimately significantly improving the predictive reliability of the 3D radio map.

[0266] Preferably, the empirical loss is implemented through the following steps: using an empirical logarithmic fitting model, generating a theoretical RSRP statistical distribution based on the physical parameters of the communication base station, and quantifying the difference between the feature vector output by the communication base station parameter encoder and the theoretical RSRP statistical distribution to constrain the large-scale path attenuation of the system.

[0267] Empirical loss is applied to the communication base station parameter encoder to ensure that the encoded base station vector representation can support the prediction results in conforming to the macroscopic path loss trend. The base station first collects its own latitude and longitude, altitude, transmit power, carrier frequency, antenna downtilt angle, and other parameters. In engineering implementation, two paths can be selected to generate a "theoretical RSRP reference map":

[0268] Classical logarithmic path loss model: Based on free space propagation or the Okumura-Hata model, the path loss is mapped to the entire voxel grid according to distance, frequency and antenna height;

[0269] Simple 3D line-of-sight propagation diagram: Directly calls the line-of-sight field intensity distribution output by the 3D propagation mechanism modeling module, and uses it as an "enhanced version" or alternative to the logarithmic model;

[0270] The generated result is a theoretical power map with the same size as the voxel mesh, reflecting the macroscopic path decay trend.

[0271] During system training, the communication base station parameter encoder has embedded the base station physical information into a high-dimensional vector tensor. Through interpolation or sampling, the vector tensor is spatially decoded into a "predicted RSRP map" with dimensions perfectly aligned with the theoretical reference map for voxel-by-voxel comparison.

[0272] Calculate the absolute difference between the predicted value and the theoretical reference value point by point, either across all voxels or only within the voxels where the prediction is the view distance. Then, spatially average all the differences to obtain a single scalar empirical loss. To avoid drastic gradients in the early stages of training, the differences can be pruned or smoothed with weights.

[0273] The gradient of the empirical loss is only propagated back to the base station parameter encoder: effectively guiding the encoder to learn the implicit rules consistent with the macroscopic path decay; avoiding other branches (image features, residual correction) from being forced to adjust synchronously, and ensuring the stability of the entire network training.

[0274] In the early stages of training, the weight of the empirical loss can be set relatively high, allowing the model to quickly grasp the general trend of "emission power decaying with distance". As the losses such as supervised regression and structural similarity gradually converge, the weight of the empirical loss can be reduced, giving the network more freedom to fit local complex scenarios.

[0275] The empirical loss ensures that the overall decay curve of the prediction results fits the classical model, avoiding distortions such as "excessive signal at the far end" or "excessive signal at the near end". The encoder has a physical prior with clear direction, so it does not learn meaningless vectors under early stochastic gradients. The macro trend is controlled by the empirical loss, while local fine-tuning is taken over by the experimental supervision, adversarial and structural loss, forming a "coarse-fine" two-level collaboration.

[0276] Preferably, the adversarial loss is implemented through the following steps: In generative adversarial training, a generator network is formed by multimodal fusion and decoder, and a discriminator network is configured. The discriminator network uses unified physical prior labels formed by fusing auxiliary radio maps for line-of-sight, waveguide effects and shadow fading with predetermined weights as real samples. Adversarial constraints are applied to the generator output through cross-entropy loss to enhance the capture of local high-frequency details and shadow effects.

[0277] The auxiliary radio map is a fusion of a line-of-sight map, a waveguide map, and a shadow fading map generated by the 3D propagation mechanism modeling module. The adversarial loss training objective is to maximize the discriminator's misclassification rate, making the generated map statistically indistinguishable from the propagation model map.

[0278] The discriminator can be a shallow 3D convolutional network. It takes a complete 3D power map as input, and the sliding convolutional kernel performs binary classification on each local voxel block, outputting a probability map to measure whether "each local region is real or generated".

[0279] The adversarial training specifically includes the following steps:

[0280] 1) Generator update: Fix the discriminator parameters; pass the batch input through the generator to obtain the prediction map;

[0281] The predicted image is fed into the discriminator to obtain the probability of it being judged as "real".

[0282] Calculate the cross-entropy to make the probability as high as possible; then add the result as an adversarial loss term to the total loss, and update the generator weights in reverse.

[0283] 2) Discriminator update

[0284] Fixed generator: First, input the prior true sample and label it as true; then input the prediction graph just generated by the generator and label it as false. Calculate the binary cross-entropy and update the discriminator to enable it to distinguish between true and false.

[0285] 3) Iterative alternation

[0286] The generator and discriminator are trained alternately at a ratio of one to one or two to one until convergence.

[0287] In the early stages of training, the weight of adversarial terms can be reduced to allow the network to learn the overall decay trend more accurately.

[0288] In the mid-stage, the adversarial weights are gradually increased to guide the network to focus on detailed textures, especially street valley stripes and building shadow edges. In the late stage, a moderate weight is maintained while ensuring the discriminator is not overly powerful, allowing for continuous enhancement of details and stable training. If necessary, spectral normalization or a small amount of noise can be added to the discriminator to mitigate gradient oscillations.

[0289] Since the generator must reproduce line-of-sight spikes, waveguide fringes, and shadow sharp edges, otherwise it would be easily detected by the discriminator, the high-frequency details of the system output are enhanced. The discriminator makes judgments at the voxel block level, prompting the generator to maintain statistical properties consistent with the physical prior in each local region. The adversarial loss mainly focuses on texture and detail. Large-scale attenuation is still controlled by empirical loss, physical prior loss, etc. The two work together to make the output conform to the overall path law and have a realistic local structure.

[0290] Preferably, the hierarchical similarity loss is implemented through the following steps: extracting feature maps of each adjacent height layer sequentially from the initial three-dimensional features output by the two-dimensional feature transfer and cross-plane mapping module; calculating and summing the squares of the element-level differences between each pair of adjacent layer feature maps to obtain the MSE between each layer; summing and normalizing the MSE between all height layers as the hierarchical similarity loss.

[0291] During the system training process:

[0292] After the initial 3D feature tensor output is completed in the 2D feature transfer and cross-plane mapping module, the tensor is split according to the height dimension to form a series of 2D feature maps, such as the 1st layer, the 2nd layer and so on up to the highest layer.

[0293] Pair layers 1 and 2, 2 and 3, and so on upwards; for each pair of adjacent layers, calculate the difference between the two feature maps at the same point for the corresponding voxel position, and then square the difference; sum up the squares of all the differences in a whole layer to obtain the "vertical smoothing error" of that layer pair;

[0294] The vertical smoothing errors of all adjacent layer pairs are summed again to obtain a total. To avoid differences in the magnitude of error due to different number of layers, the total is divided by the "number of adjacent layer pairs" or the "total number of voxels in the entire feature map" to obtain a normalized scalar, i.e., the layer similarity loss. This is then multiplied by a preset weight and added to the total loss.

[0295] The loss only backpropagates gradients to the trainable parameters of the 2D feature transfer and cross-plane mapping modules; other branches (image features, base station features, decoder) are not directly constrained by this loss, thus avoiding excessive smoothing of irrelevant weights.

[0296] In the early stages of training, the weight of hierarchical similarity loss can be increased to allow the network to establish a reasonable vertical decay pattern. As data-driven residual correction and adversarial detail loss begin to dominate local adjustments, the weight of this item can be gradually reduced to allow the network to retain the necessary height differences.

[0297] The hierarchical similarity loss constraint smooths out signal intensity variations between adjacent height layers, conforming to the common rule that "the greater the height, the smaller the path loss but the smoother the change." This significantly reduces gradient oscillations caused by random jumps in predictions between adjacent layers, thus aiding convergence. This loss requires no additional data labels, relying solely on the network's output, resulting in low computational cost. Simultaneously, it does not affect planar detail or global trend losses, complementing other physical priors and adversarial terms. The hierarchical similarity loss acts as a "vertical smoother" throughout network training, ensuring that the variation of 3D features in the height direction is both physically intuitive and avoids sacrificing necessary height differences due to excessive smoothing.

[0298] Preferably, the physical prior loss is implemented through the following steps: using the free space path loss formula, the theoretical signal power of the highest altitude layer is calculated; the difference between the average predicted value and the theoretical value of the two-dimensional feature transfer and cross-plane mapping module at this altitude layer is quantified by mean square error as the physical prior loss.

[0299] In this embodiment, based on the base station transmit power (including power after antenna gain), carrier frequency, direct distance between the base station and the center point of the highest altitude layer, and other link budget constants (system loss, receiver gain, etc., if any), a single scalar—the theoretical power value—is generated according to the classical free space path loss model. This value represents "the field strength that the transmitted signal should have when it reaches the average position of the highest altitude layer without reflection or fading".

[0300] If the scenario involves multiple base stations, the power can be calculated independently for each base station, and then the maximum value can be taken in space or the power can be superimposed to obtain a unified theoretical power; for simplified implementation, the main serving base station can also be used as a representative.

[0301] During the system training process:

[0302] In the 3D power tensor output by the 2D feature transfer and cross-plane mapping module, the layer with the largest height index is selected. A global average of the power values ​​of all voxels in this layer is then performed to obtain a scalar, called the "average value of the highest layer network". To avoid the influence of occasional outliers, a slight truncation or median filtering can be added before averaging.

[0303] The difference between the "theoretical power value" and the "average value of the highest layer network" is taken and squared. Instead of making fine-grained comparisons for individual voxels, the "overall average" is used as a single-dimensional measure. The purpose is to constrain the global trend without forcing each voxel to conform to the theory.

[0304] Multiply the squared difference by a preset weight and add it to the total loss. The weight can be adjusted according to the scene size and measurement point density: the weight can be higher when the measurement points are sparse and lowered when the measurement points are sufficient.

[0305] The physical prior loss only backpropagates the weights of "2D feature transfer and cross-plane mapping" and "3D interpolation and height correction", focusing on keeping the height attenuation curve reasonable; the detailed texture and adversarial detail branch after multimodal fusion are not directly affected by this loss, thus avoiding weakening the local high-frequency performance.

[0306] During the system training warm-up phase, a relatively large proportion of physical prior loss is allowed to quickly bring the network output to a reasonable level. In the refinement phase, the weights are reduced to a level comparable to supervised regression and structural similarity, allowing the network to retain its adaptive ability to local complex scenarios while adhering to macroscopic laws. In the final phase, a low constant weight is maintained, which serves only as a "guardrail" to prevent the highest layer prediction from going astray.

[0307] The constraint of the physical prior loss ensures that the average power of the network at the highest layer does not deviate from the free-space decay curve, fundamentally guaranteeing that the 3D map maintains reliable values ​​in high-altitude or long-distance scenes. The average value of the highest layer is like a "constant baseline," providing a fixed reference point for the entire network, making the gradient direction clearer and convergence smoother. When local adversarial and structural losses drive the network to pursue texture, the physical prior loss promptly corrects the macroscopic power level, avoiding overall underestimation or overestimation.

[0308] Preferably, the structural similarity loss is implemented through the following steps: extracting building outline masks from satellite remote sensing images and calculating the gradient map of the masks; extracting signal shadow gradient maps from the three-dimensional RSRP prediction map output by the decoding module at the corresponding height layer; calculating the structural similarity index of the signal shadow gradient map and the gradient map of the corresponding mask, and using the absolute value of the difference between it and 1 as the structural similarity loss.

[0309] The constraint process for structural similarity loss during system training includes the following steps.

[0310] 1) Obtaining building contour mask

[0311] Perform tone equalization or dehazing on high-resolution satellite images to enhance the contrast between building roofs and the background.

[0312] Using existing semantic segmentation networks (such as DeepLab, Mask2Former, and other urban target segmentation models), only the "building" category is retained to obtain a binary mask.

[0313] Perform a morphological closing operation on the binary mask to remove small holes; optionally, slightly dilate the contour to allow the mask boundary to cover the eaves shadow, avoiding negative deviations during alignment.

[0314] 2) Contour gradient map extraction

[0315] Apply the Sobel or Canny edge operator to the building mask to obtain a "building gradient map".

[0316] The gradient magnitude is linearly scaled to the 0–1 range to form a gradient intensity map, which is used for subsequent comparisons.

[0317] 3) Generation of signal shadow gradient map

[0318] For the 3D RSRP prediction tensor output by the multimodal fusion and decoding module, select one or more height slices near the ground layer (usually the layer closest to the roof height) because the shadows of these layers are most obvious.

[0319] Normalize the RSRP value first, then invert it to make the area with weak signal behind the building become a "strong response"; similarly, use the Sobel or Canny operator to obtain the "shadow gradient map"; use the same normalization method to pull it to the 0-1 interval.

[0320] 4) Calculation of structural similarity index

[0321] Using an 11×11 or 7×7 sliding window, the local mean, variance, and cross-correlation of the building gradient map and the shadow gradient map are compared position by position; these statistics are used to calculate the window-level structural similarity index (SSIM), ranging from 0 to 1.

[0322] The average SSIM of all windows is taken to obtain an overall structural similarity score; the closer the score is to 1, the more consistent the shadow edge is with the building outline.

[0323] 5) Loss definition and backpropagation

[0324] Subtracting the full-image SSIM score from 1 and taking the absolute value yields the "structural similarity loss" scalar; if multiple height layers of shadow maps are used simultaneously, their losses can be averaged.

[0325] Gradients are backpropagated only for parameters from the multimodal fusion layer to the end of the decoder; the satellite image segmentation network and the building mask themselves do not participate in backpropagation.

[0326] During training, it is recommended to keep the structural similarity loss weight at a moderate value, so that the network pays attention to both physical trends and measurement point fitting, as well as the constraints of building outlines on shadow edges. After thousands of optimization steps, it can be clearly seen that the signal shadows in the poorly lit areas such as the west and north sides of the building are successfully aligned, the outline clarity is improved, and the "light leakage" or "too dark" phenomenon in the error heatmap is significantly reduced.

[0327] Through this process, the model effectively binds the "visual building outline" with the "power shadow edge", improving the spatial alignment accuracy and visual consistency of 3D radio maps in complex scenes such as street valleys and building clusters.

[0328] Example 2

[0329] This invention provides a method for generating a three-dimensional radio map in ultra-low altitude domain. The method outputs a three-dimensional radio map characterizing the spatial distribution of radio signal power within a target area. This map simultaneously reflects the planar distribution information and vertical height information of the radio signal power in the target area. The method includes the following steps:

[0330] The target area is divided into a voxel grid containing latitude, longitude, and altitude coordinates, and sparse measured RSRP data is mapped to the corresponding voxel cells; the voxel grid includes at least a ground layer and one or more altitude layers along the vertical direction;

[0331] Convert satellite remote sensing images of the target area into high-dimensional environmental image features;

[0332] Map the physical parameters of each communication base station within the target area to the spatial propagation characteristics of the base station;

[0333] Based on the high-dimensional image features of the environment and the spatial propagation features of the base station, radio signal features of the ground layer are extracted, and the radio signal features of the ground layer are extended to each height layer based on a voxel grid to generate an initial three-dimensional feature representation.

[0334] Spatial interpolation is performed on missing voxels based on physical propagation priors and initial 3D feature representations; residual correction is applied to the interpolation results using sparse measured RSRP data, and the corrected 3D RSRP features are output.

[0335] The corrected 3D RSRP features, environmental high-dimensional image features, and base station spatial propagation features are fused together, and the fusion result is input into a 3D decoding network to generate the final 3D radio signal power distribution map.

[0336] This example uses a real-world scenario of "a core urban area of ​​10 square kilometers with a street elevation difference of no more than 120 meters" as an example. The specific reasoning process is as follows:

[0337] 1) Voxel mesh generation and measurement point placement

[0338] First, transform the latitude and longitude of the target area to the local plane coordinate system (such as UTM projection) to ensure that the Cartesian grid assumption holds true for subsequent convolution operations;

[0339] With a horizontal resolution of 5 meters × 5 meters and a height layer interval of 10 meters, a three-dimensional index of depth-height-width is automatically generated, totaling 24 layers.

[0340] Acquire measured RSRP data. The number of measured RSRP points is generally uneven, so high-density sampling is prioritized for the ground and lower atmosphere. Write each measurement point data into a grid array and mark the corresponding position as "measured", while other positions are marked as "to be completed" by default.

[0341] 2) Preparation of prior features of environment and base station

[0342] Publicly available remote sensing imagery (30 cm resolution) was used. White balance was first applied and the image was then stitched together before being fed into a multi-layer visual Transformer. The Transformer output channels contained information such as ground feature semantics, roof elevation, and shadow probability, and the resolution was perfectly aligned with the reference plane.

[0343] The base station location, transmit power, carrier frequency, antenna height, and elevation angle are converted into numerical vectors. A two-layer perceptron is used to encode these vectors into the same dimension as the image features, and finally, a plane-aligned base station propagation tensor is obtained.

[0344] 3) Ground layer feature extraction and 3D extension

[0345] A pre-trained 2D model is invoked, and the image features are concatenated with the base station propagation tensor and then input into a 2D radio map network that has been trained on a macro base station dataset. The network outputs multi-scale convolutional features at the ground layer, including information such as direct attenuation and building edge diffraction.

[0346] The parameters of the ground convolution kernel are copied to the 3D convolution kernel and convolved in parallel at each height layer. Then, hierarchical cross attention is used to correlate the results of different layers to generate an initial 3D feature tensor covering all 24 layers.

[0347] 4) Three-dimensional interpolation completion and residual fine-tuning

[0348] Based on the distance from the base station to each voxel, the height difference, and the carrier frequency, a macroscopic free space trend map is quickly generated and aligned with the initial features.

[0349] Kriging interpolation is performed only on the voxel running space of the "to be completed" voxels, taking into account both macroscopic trends and measurement point residuals during interpolation, to generate the "first version" of the full-coverage power map.

[0350] The interpolated image and the residual mask of the measurement point are stitched together and fed into a U-Net with a depth of only two levels; the U-Net outputs a voxel-level fine-tuned residual image, which is added to the interpolated image to obtain the "corrected 3D RSRP feature".

[0351] The U-Net has been trained offline. During the inference phase, it only requires one additional forward pass, and the overhead accounts for about 15% of the overall inference.

[0352] 5) Multimodal fusion and 3D decoding

[0353] The "corrected RSRP features", image features, and base station propagation features are fed into the cross-attention module; the attention output is added element by element after three-dimensional position encoding to ensure that the network perceives the absolute spatial coordinates of all voxels.

[0354] The decoder consists of four levels of three-dimensional transposed convolutional layers, with each level upsampling by 2x simultaneously in three directions;

[0355] In the upsampling stage at the depth front, high-resolution features from the ground layer are introduced to make skip connections to restore the street outline and building shadow sharpness.

[0356] Finally, a 1×1×1 convolution is used to compress the multi-channel features into a single-channel power map, which is then multiplied by a fixed scaling factor to convert it to dBm. The resulting tensor size corresponds one-to-one with the initial voxel grid and can be directly stored on disk in NetCDF or GeoTIFF format for backend business use.

[0357] After the model weights are loaded on the central server or edge nodes, each step typically takes tens of seconds to several minutes, depending on the area size, grid resolution, and hardware performance. For a typical urban area of ​​1 square kilometer, the entire inference and correction process can usually be completed within 10–30 seconds. The resulting 3D radio map can be directly applied to downstream business scenarios such as UAV flight path planning, urban network coverage blind spot identification, and base station deployment site optimization, providing a detailed and dynamic predictive basis for spatial communication analysis.

[0358] like Figure 3As shown, this embodiment focuses on the planning and optimization of urban low-altitude (0–100 meters) communication networks, providing specialized network deployment solutions for scenarios such as unmanned aerial vehicle (UAV) flight path communication and eVTOL corridors. For example, deploying this embodiment in a smart city can design continuous communication coverage for flight corridors commonly used by delivery drones: based on the city's 3D building model and airspace corridor information, the layout and parameter configuration of key base stations are determined, enabling wireless signals to form a seamless coverage grid along low-altitude flight paths.

[0359] For urban air taxi routes, this embodiment can deploy directional antennas or small base stations on the rooftops of high-rise buildings along the route, using beamforming technology to direct signal energy towards the air corridor, effectively enhancing air link quality and suppressing ground interference. In city centers, specific air corridors are often designated for drones or eVTOLs to ensure flight safety. Using 3D radio maps, the signal coverage of these corridors at high altitudes can be used to determine whether it is necessary to add rooftop base stations, adjust downtilt angles, or modify corridor routes, ensuring that aircraft maintain sufficient signal strength along their designated routes.

[0360] Compared to traditional cellular networks, which primarily serve ground users and suffer from coverage blind spots due to antenna downtilt, drones often encounter weak signal areas and frequent handovers after takeoff. With the evolution of 5G / 6G networks, many base stations are equipped with 3D beamforming capabilities (such as Massive MIMO). If only 2D maps are used for planning, frequent "over-coverage" or interference conflicts between adjacent cells may occur at high altitudes. 3D maps can help identify vertical coverage overlap between airborne and ground users, thereby optimizing antenna tilt angles, power configurations, reducing interference, and improving spectrum efficiency.

[0361] When drones or urban air vehicles fly at different altitudes, their movement paths in space form "three-dimensional trajectories." With the support of 3D radio maps, signal quality changes over time can be simulated, identifying coverage capabilities at key points along flight paths, such as sharp turns and gaps between buildings, to assess communication risks and help develop safe and reliable flight plans. Users and base stations at different altitudes may experience uplink and downlink interference on the same frequency band. 3D maps can visually display the vertical coverage "overlap" areas of each base station, allowing engineers to quickly assess which "high-altitude" areas experience the most severe interference and implement layered switching strategies or dynamic beamforming to mitigate the interference.

[0362] Two-dimensional methods often focus only on ground coverage, which can easily lead to misjudgments of the actual signal attenuation patterns in low-altitude environments, resulting in incorrect decisions regarding base station site selection or power configuration. 3D maps can more accurately predict weak or over-coverage phenomena that may occur at each altitude level before engineering deployment, significantly reducing the workload of extensive drive testing and parameter tuning afterward.

[0363] While traditional ray tracing can also calculate 3D propagation, it is usually computationally expensive and requires a long simulation time for complex urban environments. This embodiment uses a data-driven + physical prior approach to learn 3D coverage features with fewer field measurements and generate results with similar accuracy to ray tracing, but with faster inference speed and dynamic updates (such as when the configuration of urban buildings or base stations changes).

[0364] This embodiment compensates for insufficient vertical coverage through three-dimensional network planning: on the one hand, it achieves directional coverage of key low-altitude air routes, significantly reducing control link interruptions caused by weak signals or handovers during UAV flight; on the other hand, by optimizing site selection and antenna orientation, it significantly reduces co-channel interference caused by multiple base stations simultaneously radiating to airborne terminals. In actual deployment, the specially optimized low-altitude communication network can increase the signal coverage probability at typical UAV flight altitudes (e.g., 50–100 meters) from less than 90% in traditional networks to nearly 100%, ensuring the communication reliability of air traffic operations. Measurement and simulation results show that using 3D maps for low-altitude communication planning, compared to using only a two-dimensional ground coverage model, can reduce the coverage rate (i.e., the proportion of blind spots or weak coverage areas) by more than 40%, significantly improving the continuity of airborne communication links. Through vertical coverage management, base stations can finely classify cross-area coverage and uplink / downlink interference between airborne and ground users, reducing frequent handovers and bandwidth waste, and improving channel utilization by approximately 15–20%. In the planning of low-altitude networks in densely populated urban areas, if accurate design can be carried out using 3D maps in the early stages, it can avoid or reduce a large number of on-site flight tests and repeated base station modifications in the later stages, and it is estimated that the planning cycle cost can be reduced by 20-30%.

[0365] Therefore, this embodiment has good engineering feasibility, can significantly improve the level of urban airspace communication support, and meet the future demand for high-reliability and low-latency communication for drone logistics and air travel.

[0366] Example 3

[0367] The present invention provides a computer storage medium, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the above technical solution.

[0368] Example 4

[0369] The present invention also provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method described in the above technical solution.

[0370] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0371] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0372] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0373] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0374] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

[0375] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A system for generating three-dimensional radio maps in ultra-low altitude domains, characterized in that, include: A 3D mesh generation and data preprocessing module is used to divide the target area into a voxel mesh containing latitude, longitude, and altitude coordinates, and to map sparse measured RSRP data to the corresponding voxel cells; the voxel mesh includes at least a ground layer and one or more altitude layers along the vertical direction; The satellite remote sensing image feature extraction module is used to convert satellite remote sensing images of the target area into high-dimensional environmental image features; A communication base station parameter encoder is used to map the physical parameters of each communication base station in the target area into the spatial propagation characteristics of the base station. The 2D feature transfer and cross-plane mapping module is used to extract radio signal features of the ground layer by combining environmental features and base station propagation information, and to extend the radio signal features of the ground layer to each height layer based on voxel grid to generate an initial 3D feature representation. The 3D interpolation and height correction module is used to perform spatial interpolation completion on voxel units for which no measured RSRP data has been obtained, based on physical propagation priors and initial 3D feature representations; combined with sparse measured RSRP data, residual correction is performed on the interpolation results, and the corrected 3D RSRP features are output. The multimodal fusion and decoding module is used to fuse the corrected 3D RSRP features, environmental high-dimensional image features, and base station spatial propagation features. The fusion result is input into the 3D decoding network to generate the final 3D radio signal power distribution map.

2. The system according to claim 1, characterized in that, The 3D mesh generation and data preprocessing module defines the ground layer height as 0, and the ground layer and multiple height layers are distributed at equal intervals; and voxel units without radio reference signal power measurement values ​​are marked as "missing".

3. The system according to claim 1, characterized in that, The two-dimensional feature transfer and cross-plane mapping module is used for: The pre-trained 2D radio map generation model is invoked to extract multi-scale convolutional features of the ground layer, including the distribution of land cover categories, building edges, and signal attenuation trends, by taking high-dimensional environmental image features and spatial propagation features of base stations as input. The two-dimensional convolutional kernel parameters used to extract multi-scale features in the two-dimensional radio map generation model are copied and expanded into three-dimensional convolutional kernels in the vertical direction. The three-dimensional convolutional kernels are applied in parallel to each multi-scale feature map of the ground layer at the corresponding height layer to generate preliminary three-dimensional feature slices for each height layer. The initial 3D feature slices from each height layer are input into a multi-head cross-attention network, which fuses feature representations from different height layers to capture vertical correlations and output the initial 3D feature representation.

4. The system according to claim 1, characterized in that, The three-dimensional interpolation and height correction module is used for: Based on the empirical logarithmic fitting model, a theoretical RSRP trend field of the same size as the initial three-dimensional features is generated; By combining the trend field, initial three-dimensional features, and sparse measured RSRP data, missing voxels are filled using spatial interpolation methods. By combining sparse measured RSRP data, a trainable model is used to perform residual correction on the interpolation results, and the corrected three-dimensional RSRP features are output.

5. The system according to claim 1, characterized in that, The multimodal fusion and decoding module further includes: The cross-attention fusion layer uses the corrected 3D RSRP features as the query vector, and the high-dimensional environmental image features and the base station spatial propagation features as the key / value vectors. Multiple attention heads are processed in parallel to generate fused features. The 3D position encoder maps the 3D coordinates of each voxel to a spatial embedding of the same dimension as the fused feature, and adds it to the fused feature element by element, so that attention computation has absolute spatial awareness. The decoder takes the ground layer features output by the 2D feature transfer module and introduces them into the corresponding upsampling stage of the decoder through skip connections. It then upsamples and fuses the features layer by layer according to the planar resolution and height interlayer spacing of the voxel grid, and outputs the final 3D radio signal intensity distribution map.

6. The system according to claim 1, characterized in that, The system's training method includes end-to-end supervised training, utilizing the following combination of loss functions to optimize network parameters: Supervised regression loss is used to penalize the deviation between the true value label and the 3D radio signal power distribution map output by the multimodal fusion and decoding modules. Empirical loss is used to constrain the difference between the spatial propagation characteristics of the base station output by the communication base station parameter encoder and the theoretical RSRP statistical distribution; Combat loss is used to constrain high-frequency details and shading effects in the 3D radio signal power distribution map output by the multimodal fusion and decoding module; Structural similarity loss is used to constrain the alignment of the signal shadow region with the building outline of the target region in the 3D radio signal power distribution map output by the multimodal fusion and decoding module. Hierarchical similarity loss is used to constrain the smooth transition of the output of the two-dimensional feature transfer and cross-plane mapping module between adjacent height layers; The physical prior loss is used to constrain the deviation between the average predicted value of the highest-height layer output by the 2D feature transfer and cross-plane mapping module and the theoretically calculated value.

7. The system according to claim 6, characterized in that, The adversarial loss is implemented through the following steps: In generative adversarial training, a multimodal fusion module and a radio map decoder are combined to form a generator network, and a discriminator network is configured. The discriminator network uses unified physical prior labels formed by fusing auxiliary radio maps for line-of-sight, waveguide effects and shadow fading with predetermined weights as true samples. The generator output is subjected to adversarial constraints through cross-entropy loss to enhance the capture of local high-frequency details and shadow effects.

8. The system according to claim 6, characterized in that, The hierarchical similarity loss is implemented through the following steps: extracting feature maps of each adjacent height layer sequentially from the initial three-dimensional features output by the two-dimensional feature transfer and cross-plane mapping module; calculating and summing the squares of the element-level differences between each pair of adjacent layer feature maps to obtain the MSE between each layer; summing and normalizing the MSE between all height layers as the hierarchical similarity loss.

9. The system according to claim 6, characterized in that, The physical prior loss is implemented through the following steps: using the free space path loss formula, the theoretical signal power of the highest altitude layer is calculated; the difference between the average predicted value and the theoretical value of the two-dimensional feature transfer and cross-plane mapping module at this altitude layer is quantified by mean square error and used as the physical prior loss.

10. The system according to claim 6, characterized in that, The structural similarity loss is implemented through the following steps: extracting building outline masks from satellite remote sensing images and calculating the gradient map of the masks; extracting signal shadow gradient maps from the three-dimensional RSRP prediction map output by the decoding module at the corresponding height layer; calculating the structural similarity index of the signal shadow gradient map and the gradient map of the corresponding mask, and using the absolute value of the difference between the index and 1 as the structural similarity loss.

Images