Ground radio map generation system fusing satellite image and actually measured RSRP

Abstract

The invention provides a satellite image and actually measured RSRP fused ground radio map generation system, and the system comprises a satellite image feature extraction module which is used for converting a high-resolution satellite image of a target region into high-dimensional image features reflecting the ground feature type distribution of the region; the base station physical characteristic input module is used for converting the physical parameters of the communication base station in the target area into characteristic vectors reflecting electromagnetic propagation characteristics; the multi-modal fusion module is used for fusing the high-dimensional image features and the feature vectors to generate joint feature representation; the radio map decoder is used for generating a radio signal power plane distribution diagram of the target area through a decoding network based on the joint feature representation; and the ground RSRP correction module is used for correcting the radio signal power plane distribution diagram output by the radio map decoder so as to output a final radio signal power distribution result. According to the invention, the space-ground collaborative data source fusion is realized to improve the mapping precision.

Description

Technical Field

[0001] This invention belongs to the field of radio signal detection technology, specifically relating to a ground radio map generation system that integrates satellite imagery and measured RSRP. Background Technology

[0002] When available RSRP (Reference Signal Received Power) measurements are very sparse, existing two-dimensional radio mapping methods often struggle to maintain accuracy. Most such techniques heavily rely on dense terrestrial signal samples; when measurement data is insufficient, the prediction accuracy of purely data-driven interpolation or learning methods drops significantly. Traditional approaches may attempt to compensate for insufficient data by increasing the number of measurement points or employing approximate propagation models, but acquiring large amounts of measured data is costly and time-consuming, while simplified propagation models struggle to characterize signal propagation effects in complex urban environments, a problem particularly pronounced when data is scarce.

[0003] On the other hand, satellite remote sensing images provide rich environmental information (such as building layout, land cover type, terrain undulation, etc.), which can provide a reference for radio propagation modeling. However, existing technologies rarely combine these "aerial" data with limited ground measurements and make full use of them.

[0004] Furthermore, existing methods generally lack the integration of physical propagation laws as constraints into the modeling process. In the absence of such physical guidance, predictions relying solely on sparse measurement data and imagery information often fail to demonstrate sufficient physical consistency and reliability.

[0005] Therefore, when faced with sparse RSRP data input, current 2D radio map generation technology urgently needs to solve the core problem of integrating space-ground collaborative data sources and introducing propagation law constraints to improve mapping accuracy, and this problem has not yet been effectively overcome by existing technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the aforementioned background technology and provide a ground radio map generation system that integrates satellite imagery and measured RSRP, thereby achieving data source fusion between space and ground to improve mapping accuracy.

[0007] The technical solution adopted in this invention is: a terrestrial radio map generation system that integrates satellite imagery and measured RSRP, comprising: The satellite image feature extraction module is used to convert high-resolution satellite images of the target area into high-dimensional image features that reflect the distribution of land cover types in the area. The base station physical feature input module is used to convert the physical parameters of communication base stations within the target area into feature vectors that reflect electromagnetic propagation characteristics. A multimodal fusion module is used to fuse the high-dimensional image features and feature vectors to generate a joint feature representation; A radio map decoder is used to generate a radio signal power plane distribution map of a target area through a decoding network based on the joint feature representation. The decoder is used to recover large-scale attenuation trends and local high-frequency details in the joint features. The ground-based RSRP correction module maps measured RSRP data from sparse ground nodes in the target area to a corresponding spatial grid, correcting the radio signal power plane distribution map output by the radio map decoder, thereby outputting a more accurate radio signal power distribution result.

[0008] The beneficial effects of this invention are as follows: This invention is designed for sparse data scenarios. By combining satellite imagery with limited ground RSRP measurement data, it can maintain high prediction accuracy even under sparse measured sample conditions, overcoming the excessive reliance of traditional methods on dense drive test data. Based on the macroscopic ground feature distribution provided by satellite imagery and the electromagnetic propagation characteristics reflected by base station parameters, and combined with the information from sparse measurement points, the system forms a complete multi-source data input, effectively solving the problem that existing technologies cannot fully utilize "air + ground" information.

[0009] Furthermore, by explicitly including parameters such as base station altitude, this invention can more comprehensively reflect the spatial deployment characteristics of the signal transmitter, and more accurately infer large-scale signal attenuation in sparse measurement point scenarios; in urban areas with relatively complex elevation information, the base station altitude can be matched with the terrain changes in satellite images, which facilitates accurate mapping of the model when dealing with mountainous or high-rise building environments.

[0010] Furthermore, under the condition that ground RSRP data is scarce and unevenly distributed, random noise can help the model simulate more uncertain scenarios during the training phase, reducing the adverse effects of missing data or imbalanced distribution on mapping accuracy. The injection of random noise avoids the network overfitting to the existing few samples, allowing the system to maintain good adaptability to sparse measurement points in new environments.

[0011] Furthermore, this invention utilizes a visual Transformer encoder, a multilayer perceptron, and a multi-head self-attention mechanism to map multi-channel land cover category masks and base station physical feature vectors to the same vector space and perform interactive fusion. This not only preserves the global environmental information of the target area but also combines the electromagnetic propagation characteristics of the base station to achieve a comprehensive characterization of large-scale attenuation and local high-frequency details, further improving the accuracy and resolution of radio maps. The multi-head attention Transformer can simultaneously focus on land cover type features in different areas of satellite imagery and capture key location and building density information in complex urban environments during the fusion process with base station parameters.

[0012] Furthermore, given the lack of ground measurement points, this invention uses empirical loss to constrain large-scale attenuation, supervised regression loss to correct global prediction results, and adversarial loss to enhance the capture of local shadows and high-frequency details, with the three complementing each other. Although the measured data used in supervised regression loss is scarce, it can still effectively penalize the overall error of the output map, thereby significantly improving the reliability of the model in real-world scenarios.

[0013] Furthermore, this invention ensures large-scale physical consistency by comparing base station parameters with the theoretical RSRP distribution. Even in the absence of dense measured points, it ensures that the prediction results remain consistent with the classic path loss law and will not deviate significantly due to sparse data. By constraining large-scale attenuation through a logarithmic fitting model, basic physical guidance can be obtained without relying on large-scale ground drive tests, thus achieving fault tolerance of the mapping framework for sparse data.

[0014] Furthermore, in areas with insufficient ground measurement points, this invention can provide additional "reference true samples" in adversarial training by using auxiliary maps generated by ray tracing, enabling the model to learn more high-frequency details and shadow effects. Combining adversarial loss and supervised regression can take into account both global attenuation and complex propagation characteristics such as strong local occlusion and reflection, thereby improving the overall mapping accuracy in sparse scenes.

[0015] Furthermore, this invention divides the system's training process into two stages: the first stage uses a complete combination of loss functions to ensure systematic constraints on large-scale decay, local details, and overall prediction errors; the second stage freezes the model parameters from the previous stage and optimizes only the ground RSRP correction module. This two-stage strategy allows the model to first acquire comprehensive propagation prior knowledge, and then use sparse measured RSRP data to focus on local correction, balancing the model's global consistency with its adaptability to local data, and significantly improving the reliability of the final prediction results.

[0016] Furthermore, the ground RSRP correction module of the present invention adopts a dual encoder-decoder architecture to extract and decode features from the "preliminary predicted radio map" and the "sparse measured RSRP data" respectively. After fusing the features of the two by the alignment unit, prediction correction is performed. This not only preserves the large-scale propagation prior carried in the preliminary map, but also combines the local real measurement information provided by the sparse measured data, thereby significantly improving the performance of the radio signal power distribution map in terms of detail accuracy and overall consistency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system framework of the present invention; Figure 2 This is a schematic diagram of the training process of the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1 like Figure 1 As shown, the present invention provides a terrestrial radio map generation system that integrates satellite imagery and measured RSRP, comprising: The satellite image feature extraction module is used to convert high-resolution satellite images of the target area into high-dimensional image features that reflect the distribution of land cover types in the area. The base station physical feature input module is used to convert the physical parameters of communication base stations within the target area into feature vectors that reflect electromagnetic propagation characteristics. A multimodal fusion module is used to fuse the high-dimensional image features and feature vectors to generate a joint feature representation; A radio map decoder is used to generate a radio signal power plane distribution map of a target area through a decoding network based on the joint feature representation. The decoder is used to recover large-scale attenuation trends and local high-frequency details in the joint features. The ground RSRP correction module is used to map the measured RSRP data from sparse ground nodes in the target area to the corresponding spatial grid, and correct the radio signal power plane distribution map output by the radio map decoder, thereby outputting a more accurate radio signal power distribution result.

[0020] Specifically, the satellite image feature extraction module is used to extract high-dimensional features reflecting the distribution of land cover types from high-resolution satellite remote sensing images. First, the input satellite image of the target area undergoes semantic segmentation processing, classifying image pixels into different land cover types (e.g., buildings, roads, vegetation, water bodies, etc.). Based on the segmentation results, corresponding land cover masks are generated, with one or more mask layers for each land cover category to highlight the spatial distribution of different environmental elements. Then, these mask images or the segmented multi-channel images are input into a deep neural network to extract multi-level feature representations, mapping the land cover category information at each location into a feature vector of a predetermined dimension.

[0021] In other words, this module converts satellite imagery into a series of high-dimensional image feature representations, which serve as the visual feature input for subsequent fusion. The output image features contain information about the macroscopic environment within the target area, providing an environmental contextual basis for radio propagation modeling.

[0022] This module can directly use mature commercial software, or participate in system training together with other modules. During the training process of this module, real land cover category maps are used as auxiliary labels.

[0023] Specifically, the physical parameters of the communication base station include location, carrier frequency, antenna transmit power, vertical downtilt angle, and base station altitude. The input to the base station physical feature input module also includes random noise to enhance the system's diversity and robustness during training and inference.

[0024] The base station physical feature input module is used to encode the physical parameters of the communication base station to obtain a vector representation reflecting the propagation characteristics of electromagnetic waves. Its input includes various physical parameters of the base station, such as location, carrier frequency, transmit power, antenna downtilt angle, and base station altitude.

[0025] like Figure 2 As shown, the base station physical feature input module also integrates a base station parameter decoder. The base station parameter decoder is based on a diffusion model or other generative framework, and uses random noise and latent vector z to perform a "reverse denoising" or "generation" process, ultimately outputting estimated base station parameters; and generates loss by comparing with real base station parameters during training, thereby constraining the encoder's mapping process.

[0026] 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.

[0027] The reverse generation process of the base station parameter decoder includes: a denoising network based on a diffusion model (i.e., the "base station parameter decoder" in this system) 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 encoder to ensure that the restored result matches the true parameters.

[0028] At the base station parameter decoder, random noise is typically introduced in the following ways: 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; 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.

[0029] 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.

[0030] Base station parameters include at least the following basic parameters: location, carrier frequency, antenna transmit power, vertical downtilt angle, and base station height (which can be expanded as needed, such as base station latitude and longitude location, antenna type, etc., but must include at least the above four items).

[0031] The base station physical feature input module can employ a multilayer perceptron (fully connected neural network) structure to embed and nonlinearly transform the original parameters, mapping the aforementioned numerical parameters into internal feature vector representations. First, the base station parameters are normalized, for example, scaling values ​​such as transmit power and antenna height to an appropriate range. Then, these normalized parameters are fed as input to several fully connected layers, extracting high-level features layer by layer. The network's hidden layers can learn the mapping relationship between base station parameters and signal coverage patterns, such as distance attenuation trends and antenna directivity effects. In the output layer, the base station encoder generates a low-dimensional feature vector (or feature matrix) that encapsulates the factors influencing signal propagation by the base station.

[0032] During the training process of the base station 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 them in reverse. If the denoised result is still close to the original parameters, it means 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 the 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.

[0033] 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.

[0034] 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).

[0035] 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.

[0036] 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.

[0037] 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.

[0038] Specifically, the multimodal fusion module inputs the multi-channel ground feature category mask obtained by the pre-trained satellite image feature extraction module into the visual Transformer encoder, converts it into high-dimensional image features, and maps it to a predetermined vector space through an embedding layer; the feature vector output by the base station physical feature input module is processed by a multilayer perceptron and then mapped to a predetermined vector space through an embedding layer; using a multi-head self-attention Transformer structure, interactive fusion is performed on the respective mapped vectors to generate a joint feature representation that contains both global environmental information of the target area and reflects the electromagnetic propagation characteristics of the communication base station.

[0039] This module uses multimodal attention mechanisms, such as the Visual Transformer, to interactively fuse image features and base station parameter vectors. Specifically, before entering the Transformer, the data from each modality is projected onto a vector space of the same dimension: for satellite image features, the feature representation of the entire image is divided into several spatially located feature patches, and each feature patch vector is adjusted to a predetermined dimension through a linear transformation; for communication base station parameter vectors, they are mapped to vector representations of the same dimension through a fully connected layer. Then, these two types of vectors are used together as an input token sequence and fed into the Visual Transformer structure.

[0040] In the multi-head self-attention computation process of the Transformer, features from different modalities are interconnected: environmental features from satellite images provide ground feature information for each location, while base station parameter features carry global signal attenuation trends. Through multi-layer attention interactions, the Transformer enables the feature vector of each location to simultaneously incorporate information from the macroscopic environment and local electromagnetic propagation, outputting a fused joint feature representation. In short, the joint features generated by the multi-modal fusion module include both the influence of terrain environment on signal propagation and the superimposed base station transmission characteristics, laying a rich information foundation for the generation of the final radio power map.

[0041] Specifically, the radio map decoder is the decoding part of the generative network, used to map the fused joint features to generate a radio power distribution planar map (i.e., an RSRP coverage map) of the target area. The decoder can employ a method of progressively upsampling features using a convolutional neural network to restore the high-dimensional joint features output by the Transformer to geospatial-aligned two-dimensional grid data. Specifically, the decoder transforms the fused features into pixel-by-pixel signal strength predictions through multiple layers of convolution and interpolation upsampling, outputting a coverage map with a resolution matching the input satellite image.

[0042] During the generation process, the radio map decoder adjusts the overall signal strength based on the large-scale path loss trend provided by the base station parameters, while simultaneously overlaying the effects of ground cover occlusion and environmental fading reflected by satellite image features. Therefore, the decoding result can simultaneously reflect both the large-scale attenuation patterns and small-scale details of radio signal changes. Finally, the decoder outputs the predicted RSRP values ​​for each geographic grid location, forming a complete two-dimensional radio coverage map that reproduces both the macroscopic signal coverage outline and microscopic details such as shading and fading.

[0043] Preferably, the radio map decoder includes the following sub-modules: The Multi-head Attention module applies multi-head self-attention or cross-attention to the high-dimensional features of the input (often token sequences / feature maps from a multi-channel encoder or fusion module). With self-attention, each token interacts with the others to capture the global context; with cross-attention (e.g., incorporating other information), it is paired with a set of query or key features to highlight specific key regions. This module yields a reweighted set of feature vectors, allowing the decoder to pay deeper attention to key locations of signal attenuation or local shadow details.

[0044] The first MLP (Multilayer Perceptron): performs nonlinear transformation and dimensionality projection on the attention-weighted features; it typically contains two or more fully connected layers and activation functions, combined with residual connections and LayerNorm, to make the output features both expressive and training stable. Together with the attention module, it forms the "Transformer Block," generating updated token sequence features, laying the foundation for subsequent processing.

[0045] The Image to Token Attention module compares the existing 2D feature maps (image spatial representation) and token sequence representations from the current decoding stage to extract the correspondence between local spatial structure and global token information. This can be achieved by first converting the 2D feature maps (flattened or patched) into a sequence, and then performing cross-attention with the token sequence; or by using a dedicated "image→token" projection to calculate the attention weight of each token to a local image, helping the model focus on key pixels or regions. This step enhances the decoder's internal perception of image layout, enabling the decoder to map the fused tokens (base station + environmental information, etc.) to specific pixel locations, more accurately generating local details in the overlay map.

[0046] The Token to Image Attention module, following the Image to Token Attention module, forms a symmetrical relationship with the previous module. Here, each pixel / patch is used as a query, and the token sequence is used as the key / value pair. The attention weights of the image space to the tokens are calculated. A query Q is generated for each location (or patch) in the 2D feature map, and the dot product is performed with the key / value pair of the token sequence to obtain the attention distribution. This ensures that each spatial location can obtain the most relevant global or base station environment information from the token sequence. Through these two directions of attention back and forth ("image→token", "token→image"), the decoder can continuously interact between pixels and tokens, fusing multimodal high-dimensional representations with specific spatial resolution information, thereby improving the reconstruction accuracy and semantic consistency of the overlay map.

[0047] The Conv.Trans. (2×) module, following the first MLP (Multilayer Perceptron), upsamples the feature map, gradually restoring the spatial resolution. This allows the feature map, which might have originally been at a lower resolution, to be reconstructed to a higher size or the same size as the input image. Specifically, deconvolution or a combination of convolution and upsampling (such as nearest neighbor interpolation / bilinear interpolation) can be used to enlarge the feature map size. Common practices include ConvTranspose2D (kernel stride greater than 1) or PixelShuffle, which achieve a spatial size increase of ×2 or ×4. In radio map generation, if the initial feature map is flattened or downsampled in the Transformer, it needs to be "upsampled" layer by layer to restore it to the original map size (the same resolution as the satellite image) in order to output the RSRP prediction for each grid location.

[0048] The Conv. Trans. (2×) module is responsible for upsampling the previous low-resolution feature map to a medium resolution. At this point, the network has not yet generated the final full-resolution radio map, but it already has a certain degree of spatial refinement and retains relatively rich semantic information.

[0049] The second MLP block (stacked on the right), following the Token to Image Attention module, works in conjunction with the previous spatial upsampling convolution to perform channel-dimensional depth processing or feature transformation on each pixel or patch. For example, it might map the multi-channel vector after multi-head attention to a single-channel (or small-channel) RSRP value; or it might handle activation functions, channel compression, regularization, etc. Similar to the FFN in a traditional Transformer, this MLP might perform a non-linear transformation on the vector channels at each pixel location, ultimately outputting only one channel or a small number of channels (e.g., 1 representing RSRP; or multiple channels representing different frequency bands, etc.).

[0050] The Upscaling Conv module, following the second MLP, performs one or more convolutional / upsampling processes to ultimately expand the feature map to perfectly match the resolution of the input satellite image or target map. For example, from H / 4×W / 4 to H×W; it can also include a linear activation layer for outputting continuous values ​​(e.g., dBm). This allows the decoder's final output to become a radio power planar map directly mapped to a geographic coordinate grid. Users can visualize this as a 2D map or input it into subsequent networks for further processing.

[0051] The radio map decoder uses a series of sub-modules, such as multi-head attention, bidirectional image-to-token interaction, and multi-layer convolutional upsampling / MLP, to restore the fused multimodal features into a geographically resolved RSRP distribution map. This retains the global perception advantage of the Transformer and also utilizes the strengths of convolutional networks in image reconstruction and detail restoration, ultimately achieving high-precision, interpretable radio coverage map generation.

[0052] Specifically, the ground RSRP correction module includes: The first encoder is used to extract features from the preliminary radio signal power distribution map output by the radio map decoder to form a first encoded feature containing prior information about electromagnetic propagation. The second encoder is used to extract features from sparsely distributed RSRP measured data or point cloud data within the target area to form a second encoded feature that can be aligned with the first encoded feature. Alignment unit, used to align or fuse the first encoded feature and the second encoded feature in terms of spatial coordinates, resolution or network feature dimensions; The first decoder is used to decode the aligned or fused first coded features to preserve and refine the large-scale attenuation trends and local high-frequency details in the initial radio map; The second decoder is used to decode the aligned or fused second coded features to generate correction information or difference maps; The prediction unit combines the output of the first decoder with the output of the second decoder to correct the preliminary radio signal power distribution map, thereby obtaining a more accurate planar distribution result of the radio signal power in the target area.

[0053] like Figure 2 As shown, the system employs a supervised learning method during training, optimizing parameters on a large number of samples containing geographical environment, base station information, and measurement data. To comprehensively improve prediction accuracy and generation quality, three types of loss functions are designed to jointly constrain model learning during training: Empirical loss: used to constrain the output vector of the base station physical feature input module through the base station parameter decoder; Supervised regression loss: used to penalize the difference between the radio signal power plane distribution map output by the radio map decoder and the ground RSRP correction module and the true value label, in order to correct the overall prediction; Adversarial loss: used to constrain the high-frequency details and shadow effects in the target area captured by the radio map captured by the multimodal fusion module and the radio map decoder.

[0054] Specifically, the empirical loss is implemented through the following steps: using an empirical logarithmic fitting model, a theoretical RSRP statistical distribution is generated based on the physical parameters of the communication base station, and the difference between the feature vector output by the base station physical feature input module and the theoretical RSRP statistical distribution is quantified to constrain the large-scale path attenuation of the system.

[0055] Empirical loss utilizes existing propagation models in the field of communications as priors to constrain the model, guiding it to learn large-scale signal attenuation characteristics that conform to physical laws. In implementation, theoretical path loss or empirical signal strength distribution is calculated based on the physical parameters of the base station to obtain a reference radio power map representing the macroscopic attenuation trend.

[0056] For example, the expected RSRP values ​​for each location in the target area can be calculated using the free-space path loss formula or a validated empirical model to generate a theoretical radio coverage map. Subsequently, the radio map predicted by the model is compared with this theoretical coverage map, and the difference between the two is calculated as an empirical loss (e.g., using mean squared error to measure overall deviation). By minimizing this loss, the coverage map output by the model is made to closely approximate the expectations of the physical prior model in its overall shape, correcting the model's understanding of distance attenuation patterns from the early stages of training. In particular, this loss directly affects the features generated by the base station physical feature input module, ensuring that the signal strength they represent changes with distance in a trend consistent with the prior formula, thereby ensuring that the model's understanding of large-scale propagation characteristics conforms to physical reality.

[0057] Specifically, the adversarial loss is implemented through the following steps: In generative adversarial training, a generator network is formed by combining a multimodal fusion module and a radio map decoder, and a discriminator network is configured. The discriminator uses an auxiliary radio map generated by a deterministic ray tracing model as a real sample and compares it with the generator output. Adversarial constraints are imposed on the generator output through cross-entropy loss.

[0058] To further enhance the realism of the generated radio maps, an adversarial loss from a generative adversarial network (GAN) framework is introduced during training. The network, consisting of a multimodal fusion module and a radio map decoder, is treated as a generator, producing a radio power map based on the input data. A separate discriminator is trained to distinguish whether the radio power map originates from a real distribution or the generator. Here, the real distribution radio map is not obtained directly from field measurements but is approximated using high-precision ray tracing simulations. The ray tracing model uses detailed environmental and network parameters to simulate a radio overlay map containing shadow fading details, which is used as the discriminator's real sample; correspondingly, the generator's output predicted overlay map is considered a fake sample.

[0059] During training, the discriminator uses cross-entropy loss (based on binary classification) to identify ray-traced maps as real and generated maps as fake. Simultaneously, the generator aims to confuse the discriminator, optimizing its parameters to make the generated overlay maps as difficult for the discriminator to detect as real (i.e., the discriminator tends to judge them as real). Through this adversarial game, the generator continuously improves its output, gradually approaching the real ray-traced samples in terms of statistical distribution and visual features. The effect in radio maps is more realistic high-frequency details, such as more accurately depicting signal weakness areas caused by building obstruction and subtle fluctuations in signal strength with environmental changes. The adversarial loss, combined with the aforementioned empirical and supervised losses, ultimately enables the model to not only maintain numerical accuracy but also generate overlay maps with realistic textures.

[0060] Specifically, in addition to using the main label during training, this invention introduces several auxiliary labels to enhance the model's learning depth of real-world scenarios and propagation mechanisms. The "main label" typically refers to RSRP measurements in real-world scenarios or high-confidence data generated by more refined simulation methods (such as high-precision ray tracing), while the "auxiliary labels" include land cover classification information, statistical distributions generated by empirical logarithmic fitting models, and auxiliary radio maps output by deterministic ray tracing modules, etc., to provide additional constraints and prior knowledge during training at different stages or in different sub-modules.

[0061] The satellite image feature extraction module can directly use mature commercial software (such as remote sensing image processing software) to preprocess or classify high-resolution satellite images of the target area; it can also be jointly trained with other modules of this invention (such as the multimodal fusion module).

[0062] In the joint training mode, to enable the network to better extract ground feature information (such as buildings, vegetation, and water bodies) related to wireless propagation, an auxiliary loss function can be designed by comparing it with real ground feature classification maps (obtained by manual annotation or authoritative data platforms). This allows the features extracted by this module to more accurately reflect the distribution of ground features in the target area. These semantic features can better complement the physical features of the base station in the subsequent multimodal fusion process, thereby improving the ability to capture phenomena such as shadow fading and urban canyon effects.

[0063] To better constrain the understanding of large-scale path attenuation by the base station physical feature input module, this invention uses an empirical logarithmic fitting model to generate a theoretical RSRP statistical distribution, i.e., an auxiliary label, based on parameters such as base station location, carrier frequency, transmit power, antenna downtilt angle, and altitude.

[0064] During training, the vector output by the base station physical feature input module can be compared with the theoretical RSRP distribution using a difference metric (i.e., "empirical loss"), forcing the network to maintain consistency with the physical model in terms of large-scale path loss characteristics. In this way, the system acquires reasonable propagation priors at an early stage, reducing the risk of overfitting that may result from blindly relying on large-scale data for learning.

[0065] In addition to real RSRP measurement data, high-resolution deterministic ray tracing tools can be used to generate radio maps of the target area. Although such maps may still contain some idealized assumptions, they are usually quite detailed in their characterization of complex shadows and multipath effects, and can be regarded as high-confidence auxiliary labels.

[0066] In the first phase of training, the multimodal fusion module and the radio map decoder are treated as generator networks, while the discriminator network compares the "generator output" with the "auxiliary radio map generated by the ray tracing module." Through cross-entropy loss or other adversarial losses, the system achieves additional optimizations in capturing high-frequency details, shadowing effects, and multipath reflections.

[0067] Preferably, the training process of the system of the present invention can be divided into two stages, each targeting different data sources and training objectives, in order to gradually improve the overall accuracy of the radio map and its adaptability to sparse measured data.

[0068] In the first phase, supervised training is primarily conducted on the "Base Station Physical Feature Input Module," "Multimodal Fusion Module," and "Radio Map Decoder." Training data can include real RSRP measurements or simulated RSRP data generated by deterministic models (such as ray tracing models). When using real RSRP data, unobservable or missing data areas are masked to avoid incorrect supervised backpropagation of missing values.

[0069] The first phase of loss function combinations includes: Empirical loss: Based on the corresponding auxiliary labels, the empirical logarithm fitting model is used to constrain large-scale path attenuation, ensuring that the output vector of the base station physical feature input module is consistent with the statistically significant RSRP distribution; Supervised regression loss: Based on the corresponding main label, the power plane distribution map output by the radio map decoder is compared with the real label (real RSRP or simulated RSRP), and the difference is quantified and penalized. Adversarial loss: By using the training mechanism of Generative Adversarial Network (GAN), a discriminator network is introduced to compare the auxiliary radio map generated by the deterministic ray tracing model as a real sample (as an auxiliary label) with the generated result of the radio map decoder, thereby encouraging the network to capture high-frequency details such as complex environmental occlusion and shadow fading.

[0070] In the first stage, through the constraints of the aforementioned comprehensive loss function, the network can simultaneously learn: Large-scale path attenuation patterns caused by the interaction between base station physical characteristics and the ground environment; fine-grained local features such as shadow fading and high-frequency reflection; ensuring that the generated radio map approximates the real (or simulated) RSRP value at both the overall and detailed levels.

[0071] After completing the first stage, a preliminary model is obtained that can directly predict the radio signal power distribution map of the target area from "satellite image features + base station physical features". The network parameters such as the fusion module and decoder in the model reach a relatively stable state at this time and have relatively complete prior knowledge of wireless propagation.

[0072] In the second stage, the model primarily focuses on its ability to locally correct sparse measured RSRP data. Therefore, this stage only trains the encoder and decoder in the ground RSRP correction module, while freezing the parameters of the remaining modules learned in the first stage (including the base station physical feature input module, the multimodal fusion module, and the radio map decoder).

[0073] A small amount of real-world RSRP measurement data can typically be obtained in the target area, such as through mobile terminals, road test vehicles, or fixed monitoring equipment. This sparsely distributed data can provide more accurate observations for targeted correction of the preliminary radio map generated in the first phase.

[0074] The training objective of this stage is to minimize the difference between the corrected output and the real ground RSRP measurement data using supervised regression loss, without performing the adversarial training or empirical loss used in the first stage.

[0075] This training, specifically targeting the true values ​​of sparse measurement points, enhances the ability to perceive and correct local uncertainties or complex fading conditions within the region.

[0076] The parameters of the base station physical feature input module, multimodal fusion module and radio map decoder trained in the first stage remain unchanged in this stage, thereby ensuring that the large-scale environmental attenuation law and high-frequency detail capture capability can be fully preserved, and avoiding the destruction of global prior by "overfitting" of sparse measurement point data.

[0077] After this training phase, the two encoders and decoders in the calibration module achieved the best fitting capability for sparse measurement point data, and through fusion with the preliminary radio map, significantly improved the local accuracy and overall consistency of the final power plane distribution.

[0078] Compared to training only in the first stage or simple interpolation, introducing a second stage of training can make full use of real-world RSRP observations and better bridge the gap with the actual scenario.

[0079] In summary, through the two-stage training strategy described above, this invention can first learn the general laws of electromagnetic propagation and the details of complex environments in large-scale scenarios, and then perform correction and fine-tuning on a small number of real measurement points to achieve high-precision prediction of the radio signal power distribution in the target area, while taking into account both global consistency and local reliability. In the second stage, only the encoder and decoder in the ground RSRP correction module are trained, fully utilizing the correction value brought by sparse measured data while avoiding disturbance to other modules of the system, effectively improving the generalization ability and practicality of this system.

[0080] This embodiment combines the three types of losses with certain weights to construct the overall loss function of the model, and optimizes the parameters of the generator network iteratively during training. Empirical loss provides a guarantee of physical rationality, supervised loss ensures the accuracy of observation points, and adversarial loss improves the realism of the results in terms of detail. The model trained under multiple constraints can better learn the mapping relationship from multi-source inputs to radio coverage maps, and has good generalization and realism while ensuring accuracy.

[0081] This embodiment, based on the existing architecture of generating a two-dimensional radio map by "fusing satellite imagery, base station parameters, and sparse RSRP measurement data", aims to add several objective constraints during the training phase to help the model more accurately characterize the attenuation intensity of radio signals and the shape of shadow areas. Essentially, these new constraints, through multi-task or additional loss methods, work together with the original (1) empirical loss, (2) supervised regression loss, and (3) adversarial loss to form a more comprehensive objective function, enabling the trained model to have higher resolution spatial prediction capabilities.

[0082] Specifically, there are two categories of additional objectives: Attenuation amplitude constraint: forces the model to accurately reproduce the signal attenuation gradient in ray tracing or empirical formulas on the base station-user path.

[0083] Shadow boundary structure constraint: explicitly requires that the shadow distribution output by the model has a high similarity to the edge of the real (or simulated) distribution, so as to generate more realistic signal coverage in complex scenes such as building occlusion areas and deep shadow areas.

[0084] The data for attenuation magnitude constraints can be obtained from the following sources: Ray tracing curves: If ray tracing or other high-precision propagation simulation methods are used, several representative base station-user location pairs can be selected in advance to obtain real (or high-fidelity simulated) signal strength variation curves with distance. These curves can reflect typical attenuation patterns under unobstructed (line-of-sight) and obstructed (non-line-of-sight) conditions.

[0085] Empirical formula curve: RSRP values ​​calculated using empirical models such as Okumura-Hata and COST-231 at different distances can also be used to obtain a theoretical curve of attenuation as a function of distance. If reference curves for certain scenarios are known in engineering, they can also be used in combination.

[0086] In each training iteration, for a batch of samples, several "base station-user path profiles" can be randomly selected for comparison. Let M be the radio map output by the model. pred For paths (x, y), the following is an example of how to construct the attenuation magnitude loss: in: Π represents the set of paths selected in the current batch; (x d ,y d ) represents the corresponding grid coordinates in actual space when the distance to the base station (d); f ref (d) Available ray tracing results f true (d) or empirical formula f emp (d), and even a reference curve obtained by weighting them; In implementation, mean squared error (MSE) or mean absolute error (MAE) can be used.

[0087] Unlike simply comparing the MSE of the full image with a reference image, this profile-level attenuation constraint allows the model to directly focus on "how the signal strength attenuates as the distance increases"; especially when nonlinear occlusion occurs, it also allows the model to learn a more accurate attenuation gradient.

[0088] In the total loss function, the decay magnitude constraint often works synergistically with the "empirical loss": the former focuses on local gradient fitting, while the latter focuses on the global large-scale morphology, together ensuring the model's learning depth and overall consistency of the distance decay curve.

[0089] For even greater precision, the attenuation slope (such as df / dx) along the same path can be measured to form an "attenuation slope loss" to further capture the signal descent rate.

[0090] The data sources for the shadow boundary structure constraints can be obtained from the following sources: Thresholding segmentation: In the ground truth reference map (such as a ray-traced cover map), pixel regions with RSRP below a certain threshold T are marked as "shadow areas," and their contours or edges are then extracted to obtain the "shadow boundaries." Similarly, the cover map output by the model is segmented using the same thresholding method to obtain the predicted shadow boundaries.

[0091] Multi-threshold synthesis: To obtain boundary structures across different signal intensity ranges, multiple thresholds can be used to generate a series of concentric contour lines. This allows the model to learn richer shadow transition zone features.

[0092] The loss definition for shadow boundary structure can use the boundary alignment metric Hausdorff distance: evaluating the difference between two boundary curves at their maximum or average distance. In adversarial or GAN frameworks, this shadow structure loss can also synergize with adversarial losses, allowing the model to specifically enhance the generation quality of deep shadow regions.

[0093] If ray tracing results are used as ground truth, shadow boundaries under various environments / angles can be stored in advance during the simulation phase for direct comparison during training. If high-density measured data is available (which, after interpolation, forms a full-area estimate), the same method can be used to extract shadow areas.

[0094] Because the shadow boundary loss strongly drives the model to focus on the contours of low-signal areas, if the weight is too high, it may cause a slight decrease in the prediction of high-signal areas. Therefore, it usually needs to be weighed against other losses (such as supervised regression MSE) in the overall training objective.

[0095] The final loss function of the model is obtained by weighted summation of the above loss functions.

[0096] In training practice, initial training or pre-training is typically performed with a large batch size to ensure the model has basic coverage prediction capabilities. Then, the weights of attenuation and shadow boundary losses are gradually increased to refine the model's capture of local fading and shadow structures. Adversarial losses (if GANs are enabled) can also be enabled later or fine-tuned together with these two losses.

[0097] Setting additional objectives for "attenuation magnitude" and "shadow boundary structure" explicitly emphasizes the accurate characterization of signal attenuation gradients with distance and building obstruction patterns, generating more realistic and physically consistent radio maps. Building upon the original first-stage approach of "empirical priors + supervised measurement correction + adversarial training," further refining the loss allows the network to better distinguish between the learning processes of "macro-scale" and "local shadow / attenuation details," and avoids ignoring challenging local areas solely based on global MSE. The multi-task strategy also allows the model to share representations during training, enabling collaborative learning of attenuation features, shadow segmentation, and RSRP regression—achieving richer representation capabilities with the same computational resources. Specific constraints on local attenuation and shadows can compensate for information gaps caused by insufficient measured point coverage, especially in densely built-up areas or distant high-attenuation areas where measurements are difficult to obtain. Attenuation / boundary structure knowledge from ray tracing or empirical priors can still help the network learn key spatial features through the aforementioned additional task constraints.

[0098] Example 2 This invention provides a method for generating a ground radio map by fusing satellite imagery and measured RSRP data, comprising the following steps: The satellite image feature extraction module converts high-resolution satellite images of the target area into high-dimensional image features that reflect the distribution of land cover types in the area; The base station physical feature input module converts the physical parameters of communication base stations within the target area into feature vectors that reflect electromagnetic propagation characteristics; The multimodal fusion module generates a joint feature representation from the high-dimensional image features and feature vectors; Based on the joint feature representation, the radio map decoder generates a radio signal power plane distribution map of the target area through a decoding network. This decoder is used to recover the large-scale attenuation trend and local high-frequency details in the joint features. The ground RSRP correction module is used to map the measured RSRP data from sparse ground nodes in the target area to the corresponding spatial grid, and correct the radio signal power plane distribution map output by the radio map decoder, thereby outputting a more accurate radio signal power distribution result.

[0099] Specifically, the method for generating ground radio maps by fusing satellite imagery and measured RSRP includes the following steps: 1. Data preparation: Before making actual predictions, it is necessary to prepare high-resolution satellite images of the target area, physical parameters of communication base stations, and ground RSRP measurement data.

[0100] First, the acquired satellite remote sensing images undergo resolution adjustment, cropping, and possible geo-referencing to ensure that the geographical coverage of the images matches the area to be predicted, and that the coordinate system accurately corresponds to the locations of the base stations and measurement points. Configuration data of communication base stations deployed within the target area is collected, including but not limited to: base station coordinates, carrier frequency, antenna transmit power, vertical downtilt angle, and base station altitude. Numerical parameters are normalized or range-limited, and a small amount of random noise is added if necessary to enhance robustness. This data is then passed to the base station physical feature input module as input.

[0101] Within the target area, several ground-based RSRP measurement points can be collected based on actual conditions. Each point's geographic coordinates and signal strength (e.g., RSRP value in dBm units) are recorded. If the measurement distribution is sparse, only a few key areas (such as densely built-up areas or edge coverage areas) are needed to provide a robust correction for the model. After necessary denoising and normalization processing, the measurement data is mapped into embedded features by the ground-based RSRP correction module, serving as a true reference for subsequent correction of local signal strength.

[0102] 2. Model Reasoning: Image feature extraction: The preprocessed multi-channel satellite image is input into the satellite image feature extraction module, which outputs a high-dimensional image feature tensor that reflects the land cover type.

[0103] Base station parameter encoding: Input the physical parameters (including random noise) of the base station into the base station physical feature input module to obtain a vector representation of the base station's transmission characteristics.

[0104] RSRP data integration: If several RSRP measurement points and their coordinates are known, they are input into the ground RSRP correction module to generate a sparse RSRP embedding feature map or several discrete measurement vectors.

[0105] Multimodal fusion: Image features and base station feature vectors are fed into a multimodal fusion module (using Transformer or other attention structures) to generate a joint feature representation.

[0106] Radio map decoding: The above joint features are input into the radio map decoder, which decodes and outputs an RSRP coverage prediction map of the entire target area. This prediction map is presented in the form of a two-dimensional grid, with each grid corresponding to the signal strength value at that location, thus visually demonstrating the spatial coverage of the radio network.

[0107] 3. Fine-tuning and correction To further improve prediction accuracy in specific target areas, this system supports fine-tuning or post-correction via the ground RSRP correction module after the radio map decoder outputs the preliminary radio signal power plane distribution map. The specific steps include: Acquire new or existing ground RSRP measurement point data in the target area and input them into the calibration module in the form of sparse grid or point cloud.

[0108] The calibration module has a second encoder that can preprocess, normalize, and encode the measured coordinates and signal values ​​to generate a high-dimensional representation that can be aligned with the features of the preliminary radio map. It can also perform interpolation or masking operations on the measurement data as needed.

[0109] The correction module first obtains the preliminary radio map features extracted by the first encoder, which include the large-scale attenuation prior and local environmental semantics learned in the first stage of training.

[0110] The alignment unit aligns the measurement data features output by the second encoder with the preliminary map features output by the first encoder to the same coordinate system or feature space.

[0111] In the aligned space, if it is necessary to obtain the error distribution, the “preliminary map prediction value and the measured value” can be directly compared at the measurement coordinates, or the sparse measurement information can be interpolated into a distribution map with the same resolution as the preliminary map, and pixel-level difference can be made with the preliminary map to form a power difference map / error map.

[0112] The correction module includes a first decoder and a second decoder, which process the aligned "preliminary map features" and "sparse measured features" respectively to generate or update local correction information: The first decoder retains and refines the large-scale decay trends and environmental priors in the initial map; After fusing the features of the measured values, the second decoder outputs a correction map or "correction patch map" for local deviations.

[0113] The results from the two decoders are fused within the prediction unit: higher weights are given to areas near the measurement point to approximate the true measurement value; more reliance is placed on the preliminary prediction for areas farther from the measurement point; if a large error is detected in a local area, a more refined correction is made for that area (such as using a "dense patch" strategy).

[0114] This process is performed only after the initial radio map output, without requiring modification or large-scale updates to the network weights of the radio map decoder, thus minimizing disruption to the original global prior knowledge of the network.

[0115] By fusing the outputs of the first and second decoders, the correction module can generate a new radio map, "overlaying" or "replacing" local correction values ​​into the original prediction grid, thereby obtaining a power plane distribution result that is closer to the measured distribution.

[0116] For areas without measurement coverage, the original global prior trend is maintained, avoiding gaps caused by over-reliance on sparse measurement point interpolation.

[0117] If more measurement data is subsequently obtained, and the model needs to adapt to environmental or network changes at the underlying parameter level, the encoder and decoder in the calibration module can be unfrozen and a small-scale retraining can be performed. However, if only a one-time correction is required, the above calibration process can meet most coverage correction needs without changing the other parameters of the network.

[0118] Under the two-stage training framework, the second stage can also be performed on supervised regression training only for the calibration module, thereby further optimizing its learning ability for sparse measurement points without modifying the parameters of other modules (multimodal fusion, radio map decoder, etc.).

[0119] Through the above steps, the ground-based RSRP correction module fully utilizes the large-scale propagation priors provided by the preliminary map and combines the information increments from sparse measured data to achieve a balance between "local fine-tuning" and "global consistency." Without disrupting the existing network structure and parameters, it significantly improves the accuracy of coverage prediction and the reliability of signal distribution in the target area. This embodiment is suitable for wireless coverage mapping needs in remote or difficult-to-measure areas. When the target area lacks sufficient signal measurement data, this embodiment introduces satellite imagery to obtain surface environmental features (such as village and town distribution, terrain undulations, vegetation cover, etc.) and combines them with a small number of collected base station signal samples to construct a predictive model for regional wireless coverage. In engineering applications, publicly available high-resolution satellite images can be obtained first to extract ground feature distribution information. Then, machine learning algorithms can be used to map these environmental features into the wireless signal propagation model, thereby generating a coverage intensity map for the area. Compared to traditional methods that can only assume idealized coverage when data is scarce (e.g., simply approximating the coverage area with Voronoi cells, ignoring signal holes and overlaps), this embodiment significantly improves the precision and reliability of coverage prediction.

[0120] Actual testing shows that this embodiment reduces the mean square error of signal strength prediction to about 2–3 dB, which is nearly 50% less than the error of traditional empirical models (which usually have an error of more than 5 dB).

[0121] Meanwhile, since satellite imagery is easy to acquire and has a wide coverage, this embodiment reduces the need for large-scale on-site surveys, thereby lowering the deployment costs and time investment for network planning in remote areas.

[0122] In summary, this embodiment demonstrates high engineering feasibility in communication coverage assessment in underdeveloped areas, enabling the rapid generation of reliable coverage maps and providing a basis for operators to optimize network deployment under conditions of scarce measurement data.

[0123] Example 3 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.

[0124] Example 4 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.

[0125] 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.

[0126] 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.

[0127] 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.

[0128] 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.

[0129] 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.

[0130] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A terrestrial radio map generation system that integrates satellite imagery and measured RSRP, characterized in that: include: The satellite image feature extraction module is used to convert high-resolution satellite images of the target area into high-dimensional image features that reflect the distribution of land cover types in the area. The base station physical feature input module is used to convert the physical parameters of communication base stations within the target area into feature vectors that reflect electromagnetic propagation characteristics. A multimodal fusion module is used to fuse the high-dimensional image features and feature vectors to generate a joint feature representation; A radio map decoder is used to generate a radio signal power plane distribution map of a target area through a decoding network based on the joint feature representation. The decoder is used to recover large-scale attenuation trends and local high-frequency details in the joint features. The ground RSRP correction module is used to map the measured RSRP data from sparse ground nodes in the target area to the corresponding spatial grid, correct the radio signal power plane distribution map output by the radio map decoder, and thus output the final radio signal power distribution result.

2. The system according to claim 1, characterized in that: The physical parameters of a communication base station include its location, carrier frequency, antenna transmit power, vertical downtilt angle, and / or base station altitude.

3. The system according to claim 1, characterized in that: The input to the base station physical feature input module also includes random noise to enhance the diversity and robustness of the system during training and inference.

4. The system according to claim 1, characterized in that, The multimodal fusion module inputs the multi-channel ground feature category mask obtained by the pre-trained satellite image feature extraction module into the visual Transformer encoder, converts it into high-dimensional image features, and maps it to a predetermined vector space through an embedding layer; the feature vector output by the base station physical feature input module is processed by a multilayer perceptron and then mapped to a predetermined vector space through an embedding layer; using a multi-head self-attention Transformer structure, interactive fusion is performed on the respective mapped vectors to generate a joint feature representation that contains both global environmental information of the target area and reflects the electromagnetic propagation characteristics of the communication base station.

5. The system according to claim 1, characterized in that, The system is trained using a supervised learning method, and the following combination of loss functions is constructed: Empirical loss: Large-scale path attenuation used to constrain the output vector of the base station physical feature input module; Supervised regression loss: used to penalize the difference between the radio signal power plane distribution map output by the radio map decoder and the ground RSRP correction module and the true value label, in order to correct the overall prediction; Adversarial loss: used to constrain the high-frequency details and shadow effects in the target area captured by the radio map captured by the multimodal fusion module and the radio map decoder.

6. The system according to claim 5, characterized in that, The empirical loss is implemented through the following steps: using an empirical logarithmic fitting model, a theoretical RSRP statistical distribution is generated based on the physical parameters of the communication base station, and the difference between the feature vector output by the base station physical feature input module and the theoretical RSRP statistical distribution is quantified to constrain the large-scale path attenuation of the system.

7. The system according to claim 5, characterized in that, The adversarial loss is implemented through the following steps: In generative adversarial training, a generator network is formed by combining a multimodal fusion module and a radio map decoder, and a discriminator network is configured. The discriminator uses an auxiliary radio map generated by a deterministic ray tracing model as a real sample and compares it with the generator output. An adversarial constraint is applied to the generator output through cross-entropy loss.

8. The system according to claim 5, characterized in that, The training process of the system includes the following two stages: Phase 1: Supervised training is performed based on real or deterministic RSRP data, applying all the loss function combinations mentioned above, and simultaneously training the base station physical feature input module, multimodal fusion module, and radio map decoder. If there are unobservable regions in the actual RSRP data, then the regions are masked. Second stage: Keeping the model parameters obtained in the first stage unchanged, only apply the supervised regression loss function and train the ground RSRP correction module using sparse ground RSRP measured data.

9. The system according to claim 1, characterized in that, The ground RSRP correction module includes: The first encoder is used to extract features from the preliminary radio signal power distribution map output by the radio map decoder to form a first encoded feature containing prior information about electromagnetic propagation. The second encoder is used to extract features from sparsely distributed RSRP measured data or point cloud data within the target area to form a second encoded feature that can be aligned with the first encoded feature. Alignment unit, used to align or fuse the first encoded feature and the second encoded feature in terms of spatial coordinates, resolution or network feature dimensions; The first decoder is used to decode the aligned or fused first coded features to preserve and refine the large-scale attenuation trends and local high-frequency details in the initial radio map; The second decoder is used to decode the aligned or fused second coded features to generate correction information or difference maps; The prediction unit combines the output of the first decoder with the output of the second decoder to correct the preliminary radio signal power distribution map, thereby obtaining a more accurate planar distribution result of the radio signal power in the target area.

10. A method for generating a terrestrial radio map based on the system described in any one of claims 1-9, by fusing satellite imagery and measured RSRP, characterized in that, Includes the following steps: The satellite image feature extraction module converts high-resolution satellite images of the target area into high-dimensional image features that reflect the distribution of land cover types in the area; The base station physical feature input module converts the physical parameters of communication base stations within the target area into feature vectors that reflect electromagnetic propagation characteristics; The multimodal fusion module fuses the high-dimensional image features and feature vectors to generate a joint feature representation; Based on the joint feature representation, the radio map decoder generates a radio signal power plane distribution map of the target area through a decoding network. This decoder is used to recover the large-scale attenuation trend and local high-frequency details in the joint features. The ground RSRP correction module is used to map the measured RSRP data from sparse ground nodes in the target area to the corresponding spatial grid, and correct the radio signal power plane distribution map output by the radio map decoder, thereby outputting a more accurate radio signal power distribution result.

Images