A high-precision radio environment map reconstruction method based on multi-modal fusion

Abstract

The application relates to the technical field of radio communication, and particularly discloses a high-precision radio environment map reconstruction method based on multi-modal fusion, which comprises the following steps: S01, a plurality of parallel encoders are used to process a plurality of input information respectively, and a plurality of corresponding potential feature representations are generated; S02, a fusion attention module is used to perform cross-modal dynamic fusion on the plurality of potential features to obtain a fused potential feature representation, which comprises the following steps: a spatial adaptive attention graph is constructed, a double attention mechanism is used to model spatial and channel dependencies simultaneously, the spatial adaptive attention graph is grouped, processing is performed through a spatial or channel attention sub-branch, and shuffling and connection are finally performed to promote cross-channel interaction; and S03, a decoder is used to reconstruct the fused potential feature representation, which comprises the following steps: high-resolution spatial features are reserved through a skip connection, and position information is transferred from the encoder to the decoder, so that a high-precision radio environment map is obtained.

Description

Technical Field

[0001] This invention relates to the field of radio communication technology, and more specifically to a high-precision radio environment map reconstruction method based on multimodal fusion. Background Technology

[0002] As wireless communication systems evolve towards full-scale 5G commercialization and 6G pre-research, the high-precision construction of Radio Environment Maps (REMs) has become a core foundation for achieving dynamic spectrum resource management, intelligent interference suppression, and autonomous network optimization. In practical application scenarios such as UAV-assisted networks, vehicular communications, and dynamic cognitive radio systems, the availability of reliable REMs determines the effectiveness of critical operations such as spectrum switching, load balancing, and adaptive deployment. However, due to high costs, long time commitments, and infrastructure requirements associated with intensive field measurements, especially in large-scale or dynamically changing environments, constructing REMs with fine spatial resolution is challenging. Traditional REM reconstruction methods mainly rely on interpolation algorithms, which are computationally efficient, but their effectiveness rapidly declines in heterogeneous environments characterized by complex propagation phenomena such as shadowing, diffraction, and multipath scattering.

[0003] Deep learning techniques have become a powerful tool for REM estimation. These methods often incorporate auxiliary information such as building layout, terrain elevation, or transmitter location to improve generalization and robustness.

[0004] Synthetic Aperture Radar (SAR) can capture high-resolution surface and structural information, unaffected by weather or lighting conditions, making it ideal for creating consistent environmental maps. However, SAR imagery is rarely incorporated into data-driven REM (Real Environment Model) constructions due to the following technical challenges: significant differences exist between RF measurements and SAR imagery in terms of spatial features, scale, and statistical properties, making feature alignment and fusion tasks difficult; noise, speckle, buoyancy, and shadow effects in SAR imagery can introduce irrelevant or misleading features, which, if not handled properly, can degrade model performance. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a high-precision radio environment map reconstruction method based on multimodal fusion. A Cross-Modal Complete Tensor Network (CMCTNet) is provided, designed to reconstruct a high-fidelity REM under sparse measurement conditions by fusing multimodal information from RSRP, synthetic aperture radar imagery, and building information data. It integrates three dedicated encoders for each mode, a fusion attention module (FAM) that adaptively captures cross-modal dependencies, and a skip-connection decoding pipeline for fine-grained signal recovery in complex environments.

[0006] To achieve the purpose of the present application, the technical scheme of the present application is: a high-precision radio environment map reconstruction method based on multi-modal fusion, comprising:

[0007] Step S01: a plurality of parallel encoders are used to process a plurality of input information respectively to generate a plurality of corresponding latent feature representations;

[0008] Step S02: a fusion attention module is used to perform cross-modal dynamic fusion on the plurality of latent features to obtain a fused latent feature representation, comprising: constructing a spatial adaptive attention graph, and using a double attention mechanism to model spatial and channel dependencies simultaneously, grouping the spatial adaptive attention graph, processing through a spatial or channel attention sub-branch, and finally shuffling and connecting to facilitate cross-channel interaction;

[0009] Step S03: a decoder is used to reconstruct the fused latent feature representation, comprising preserving spatial features with high resolution by jump connection, and transferring position information from the encoder to the decoder to obtain a high-precision radio environment map.

[0010] Preferably, in step S01, the plurality of parallel encoders includes three parallel encoders. The use of three parallel encoders enables modal-specific feature extraction and cross-modal collaborative enhancement: through independent RSRP, synthetic aperture radar image and building information data encoding paths, the model can respectively retain the propagation characteristics of wireless signals, the scattering characteristics of ground objects and the geometric constraints of buildings, avoiding feature confusion caused by shared encoders.

[0011] Preferably, in step S01, the plurality of input information includes sparse RSRP maps, SAR images and building data.

[0012] Preferably, in step S01, the processing of the plurality of input information includes: using discrete wavelet transform to perform initial feature extraction on the plurality of input information, decomposing the input information into a plurality of frequency bands, and performing multi-level feature extraction processing on the plurality of frequency bands through separate convolution layers respectively to generate latent feature representations.

[0013] Preferably, the plurality of frequency bands include LL, LH, HL and HH. These four frequency bands represent different frequency information of the input features respectively: the LL channel captures the low-frequency structural features of the image, mainly containing overall shape information such as contours, backgrounds, etc.; the LH channel responds to high-frequency changes in the horizontal direction, and can highlight horizontal edge features; the HL channel reflects the vertical direction edges or textures; and the HH channel concentrates on preserving the high-frequency details in the diagonal direction of the image, such as sharp corners, diagonal boundaries, etc. This frequency division mechanism enables the model to distinguish information of different frequencies at the shallow stage, thereby improving the understanding ability of structures and edges. In this way, CMCTNet can fully exploit the hidden spatial details and structural clues in the input at an early stage of the network, laying a more robust foundation for subsequent multi-modal feature extraction and fusion, especially in sparse signal recovery tasks, which can enhance the model's ability to retain key texture and edge information.

[0014] Preferably, in step S02, the cross-modal dynamic fusion is performed by taking the latent feature representation extracted from the sparse RSRP map as the query Q, and taking the fused latent feature representation extracted from the SAR image and the building data as the key K' and the value V.

[0015] Preferably, in step S02, the spatial adaptive attention map is constructed based on the following attention calculation formula:

[0016]

[0017] In the formula, Q is the RSRP query tensor, K' and V are the key and value of modal fusion, d k is the dimension of the key vector.

[0018] Preferably, in step S03, the decoder outputs a dense prediction map for estimating the complete RSRP distribution within the discrete region;

[0019] The dense prediction map is: wherein, is the complete RSRP distribution estimated by the multi-modal deep learning framework, N x and N y represent the number of grid units along the x direction and the y direction after discretization of the target region, respectively.

[0020] Preferably, each encoder includes a plurality of dense blocks. Through the layer-by-layer dense connection structure, each convolutional layer can directly access the feature maps of all previous layers, realizing the cascaded reuse of the original signal propagation mode and high-order abstract features, thereby significantly improving the modeling ability of the network for the spatial variation of RSRP, which is also the case in synthetic aperture radar images and building information data.

[0021] Preferably, the internal forward propagation process of the dense block comprises: firstly performing a first convolution on the input to obtain a feature x1, then splicing x1 and the original input in the channel dimension to obtain x2 as the input of a second convolution block; and splicing x1, the original input and x2 in the channel dimension to obtain a final output x3 as the input of a third convolution block.

[0022] The above technical solution has the following beneficial effects:

[0023] The high-precision radio environment map reconstruction method based on multi-modal fusion provided by the application introduces synthetic aperture radar images as a new mode into the data-driven REM construction, provides a robust structure prior, thereby improving the propagation modeling in challenging environments; through a unified multi-modal deep learning framework, the RSRP samples, SAR images and building data are effectively integrated through a multi-attention mechanism, thereby learning the obstacle perception signal pattern; a new benchmark dataset is used to match the synthesized RSRP map with the co-registered synthetic aperture radar image and the building information map derived from the geographic information system, thereby realizing repeatable research and comparative evaluation. Through simulation experiments, it is verified that the performance of the CMCTNet is obviously better than that of the traditional interpolation method and the latest DL-based baseline method, especially under extreme sampling sparsity and non-uniform measurement conditions. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0025] Figure 1 The flow chart of the high-precision radio environment map reconstruction method based on multi-modal fusion provided by an embodiment of the application is provided.

[0026] Figure 2 The paradigm structure diagram of the encoder-fusion-decoder provided by an embodiment of the application is provided.

[0027] Figure 3 The internal structure schematic diagram of the dense block provided by an embodiment of the application is provided.

[0028] Figure 4 The architecture structure schematic diagram of the mechanism provided by an embodiment of the application is provided.

[0029] Figure 5 The visual comparison diagram of the REM reconstructed under different sampling levels provided by an embodiment of the application is provided. DETAILED DESCRIPTION

[0030] The embodiments of the present application will be described in further detail below, obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0031] The terms "first", "second", and the like (if any) in the description and claims are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated (if any) or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units, is not necessarily limited to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0033] Embodiment one

[0034] One embodiment of the present application provides a high-precision radio environment map reconstruction method based on multi-modal fusion, which is used to reconstruct a high-fidelity REM from sparse-sampled RSRP measurements. The structural prior extracted from SAR images and building information maps is effectively combined to improve signal recovery under extreme data sparsity conditions. Learning spatially consistent cross-modal representations while considering urban propagation conditions, material-induced attenuation, and structural occlusions.

[0035] A high-precision radio environment map reconstruction method based on multi-modal fusion is as follows Figure 1As shown, the process includes: Step S01: Processing multiple input information using multiple parallel encoders to generate corresponding multiple latent feature representations; Step S02: Using a fusion attention module to dynamically fuse the multiple latent features across modalities to obtain a fused latent feature representation, including: constructing a spatial adaptive attention map and using a dual attention mechanism to simultaneously model spatial and channel dependencies, grouping the spatial adaptive attention map, processing it through spatial or channel attention sub-branches, and finally shuffling and connecting it to promote cross-channel interaction; Step S03: Reconstructing the fused latent feature representation using a decoder, including preserving high-resolution spatial features through skip connections and transferring location information from the encoder to the decoder to obtain a high-precision radio environment map.

[0036] CMCTNet is a multimodal deep learning network designed for high-fidelity radio environment map (REM) reconstruction tasks, such as... Figure 2 As shown, the overall architecture adopts an "encoder-fusion-decoder" framework, aiming to recover a complete electromagnetic signal distribution map from sparse RSRP measurements. The network first receives three input modalities: a sparse RSRP map, a SAR remote sensing image, and a building information map, and extracts features through three independently designed modal encoders. Each encoder front-end embeds wavelet blocks to achieve multi-scale frequency domain modeling, enhancing the ability to recognize texture edges, structural contours, and spatial details. Subsequently, each modal feature is further processed by dense block extraction to extract deep feature representations. After encoding, the three modal features are fed into a fusion attention module. This module utilizes a self-attention mechanism, using RSRP features as queries and SAR and building information as keys, and dynamically guides the fusion of cross-modal features through channel-spatial joint attention, thereby modeling the propagation characteristics of occlusion perception. The fused unified features are then fed into the decoder for progressive upsampling to restore the original spatial resolution. Simultaneously, shallow features from the encoder are introduced through skip connections to preserve key signal boundaries and local variation information. Finally, the network outputs a fully reconstructed RSRP distribution map. CMCTNet utilizes multimodal feature co-enhancement to demonstrate superior completion and structure preservation capabilities in low sampling rates and complex urban scenarios, providing an efficient and robust solution for radio environment perception under sparse sensing conditions.

[0037] In step S01, the plurality of parallel encoders includes three parallel encoders. The plurality of input information includes sparse RSRP maps, SAR images and building data. Processing the plurality of input information respectively includes: performing initial feature extraction on the plurality of input information by using discrete wavelet transform, decomposing the input information into a plurality of frequency bands, and performing multi-level feature extraction processing on the plurality of frequency bands respectively by a separate convolution layer to generate latent feature representation. In this embodiment, in order to enhance the modeling ability of the network to local signal discontinuity and structural patterns, a wavelet-based deep separable convolution module is constructed, which is one of the core structures of each modality encoder and is embedded in the front end of the network to improve the feature extraction effect. Unlike the traditional convolution method that only operates in the spatial domain, this module combines multi-layer discrete wavelet transform (DWT) and inverse wavelet transform (IWT) to realize frequency domain enhanced feature representation. First, the input feature map is decomposed by DWT to generate four subbands: LL, LH, HL and HH, which correspond to different directions and frequency components. Then, the four subbands are processed using a special channel grouping convolution, and the feature amplitude is further adjusted by a learnable scaling module to enhance the expression ability of building edges, signal occlusion and multipath reflection in high-frequency components. During multi-layer wavelet decomposition, the LL component extracted by each layer is retained and passed to the next layer for further decomposition, while the processed high-frequency components (LH, HL, HH) are reconstructed with the current LL in the decoding stage, and the resolution scale is restored by IWT. This encoding-decoding type wavelet processing framework enables the module to aggregate multi-scale feature information while maintaining consistent spatial resolution. At the same time, the module also designs a parallel spatial domain convolution branch to extract shallow local features, and element-wise weighted fusion is performed with the reconstructed wavelet features to integrate the advantages of frequency domain and spatial domain. In addition, the module supports optional spatial downsampling operation, and maintains channel independence through grouped convolution to avoid information aliasing.

[0038] To enhance the expression ability and information flow efficiency of the feature extraction module, a dense block is designed and introduced in the encoder. This module draws on the dense connection idea of DenseNet, and uses cascading feature fusion mechanism to effectively alleviate the problems of gradient disappearance and feature redundancy in deep network. The internal structure of the dense block is as follows Figure 3As shown, in the forward propagation, first, the input is subjected to the first convolution to obtain the feature x1, and then x1 is spliced with the original input in the channel dimension to obtain x2 as the input of the second convolution block. Then, x1, the original input and x2 are spliced in the channel dimension to obtain the final output x3 as the input of the third convolution block. In this embodiment, the dense block contains three convolution units in series, and the input of each layer of convolution operation includes not only the output of the previous layer, but also the original input feature, so as to realize cross-layer feature fusion and reuse. First, the input tensor is processed by the first convolution unit to generate the first stage feature; the feature is spliced with the input tensor and sent to the second convolution unit, and the output is further enhanced; then, the second stage output is spliced with the previous two features again and sent to the third convolution unit, and finally the fused output feature is obtained. Through such multiple splicing and convolution operations, the dense block can effectively capture local spatial details and cross-scale structural features, and is especially suitable for modeling non-stationary electromagnetic propagation patterns in complex urban environments. In order to construct the basic convolution unit used in this module, a lightweight convolution encapsulation module Conv2D is designed, which is composed of a 3x3 standard convolution layer, a batch normalization (Batch Normalization) layer and a ReLU activation function in series, for stabilizing training and enhancing nonlinear expression ability. This structure not only maintains the spatial consistency of the features, but also has good parameter efficiency and computational stability.

[0039] After encoding, the extracted specific modal features are aligned and fused through the fusion attention module (FAM), which is the core of the CMCTNet cross-modal reasoning capability. This module adopts a scaled dot-product attention mechanism to dynamically model the interaction between the RSRP signal features (as queries) and the environmental priors extracted from SAR and building data (as keys and values). In order to inject building geometry information into the attention process, the building features are added to the SAR key tensor in an element-wise manner, forming K' = K SAR +X B The calculation formula of the attention output is:

[0040]

[0041] In the formula, Q is the RSRP query tensor, K' and V are the keys and values of modal fusion, d k is the dimension of the key vector. This formula enables the model to construct a spatially adaptive attention map, taking into account geometric obstacles, material-dependent scattering, and other propagation-related interactions.

[0042] To further enhance the representation capability of the model in the multi-modal information fusion process, a double attention mechanism module is introduced, such as Figure 4As shown, the channel attention (Channel Attention) is combined with the spatial attention (Spatial Attention), and assisted by the channel shuffle operation, which effectively enhances the diversity and discriminability of feature expression. In the embodiment, the double attention mechanism module first reorganizes the input features according to the specified grouping number G, so that each channel subset can calculate the channel attention and spatial attention respectively. For the channel attention path, the module first performs global average pooling on the input features to extract channel-level statistical features, then adjusts the channel response strength through the learnable scaling and bias parameters, and finally obtains the channel weight through the Sigmoid activation function to realize the strengthening of the key signal dimension. The spatial attention path adopts GroupNorm to normalize the spatial dimension, retain the position-sensitive features, and then highlight the region response with spatial distinction through the same weight bias adjustment and activation operation. The outputs of the two paths are spliced and fused in the channel dimension, and finally the original order is scattered through the channel shuffle operation to promote information exchange and feature mixing between different channel groups, thereby effectively alleviating the redundancy problem between attention channels. The overall structure of the module is lightweight, the parameters are learnable, and the modeling capability of the network for complex structures, edge transitions and inter-modal coupling features can be significantly enhanced while keeping the computational overhead low.

[0043] The decoder component of CMCTNet reconstructs the complete RSRP distribution from the fused latent representation. Its structure is similar to the RSRP encoder and introduces skip connections to preserve high-resolution spatial features throughout the network. These connections directly transfer location information from the encoder to the decoder, which is particularly important when reconstructing fine-grained signal variations that can be lost at deeper levels. The output of the decoder is a dense prediction map: for estimating the complete RSRP distribution within a discrete region. Where, is the estimated complete RSRP distribution by the multi-modal deep learning framework, N x and N y represent the number of grid cells along the x-direction and y-direction, respectively, after discretizing the target region.

[0044] The REM reconstruction problem can be formulated as a tensor completion task, which aims to reconstruct the complete discretized RSRP distribution from sparse measurements and related SAR images. Consider a two-dimensional region of interest A with length X and width Y. This region is discretized into a uniform rectangular grid G, with size G ∈ where each grid cell has a length of Δx and a width of Δy, satisfying the following relationship:

[0045]

[0046] where (N x ) and (N y ) denote the number of grid cells along the x and y directions, respectively, as shown in Figure 1 The spatial coordinates associated with each grid cell ((i,j)) are defined as:

[0047] s(i,j) = (i · Δ x , j · Δ y )

[0048] where i ∈ {1,2,…,N x}, j ∈ {1,2,…,N y}. In this spatial discretization model, the signal power at each grid location is represented by a function (P(s,f)) that maps coordinate-frequency pairs ((s,f)) to the corresponding RSRP measurements:

[0049] P(s(i,j),f) = RSRP at s(i,j) for frequency f

[0050] Define the fully observed but unknown RSRP distribution as a tensor:

[0051] Y ∈ R Nx×Ny ,

[0052] where each element (Y (i,j) ) represents the RSRP value at grid cell ((i,j)). Due to practical limitations, only a finite number of RSRP measurements are available, leading to an incomplete representation of the signal distribution.

[0053] To improve the accuracy of RSRP reconstruction and mitigate the effects of missing data, incorporate additional environmental information from remote sensing data, specifically SAR imagery and building maps. SAR imagery provides information on structures and surface roughness that affect radio signal propagation. Define the SAR-derived feature tensor as: X SAR ∈ R Nx×Ny .

[0054] Similarly, the building information map, which captures the impact of urban structures on radio frequency propagation, is represented as: X B ∈ R Nx×Ny .

[0055] Given the sparse RSRP observations Y OBS and auxiliary data X SAR and X B , estimate the complete RSRP distribution

[0056]

[0057] where f(·) denotes the proposed multi-modal learning framework. This problem setting enables the multi-modal learning framework to exploit the spatial correlation between radio frequency signal measurements and environmental priors. In particular, SAR imagery and building information maps provide complementary cues about propagation obstructions, surface discontinuities, and structural layouts, which significantly influence signal attenuation and coverage. To train and evaluate the proposed multi-modal REM reconstruction framework, a dataset is constructed that integrates simulated RSRP measurements, co-registered SAR imagery, and building information maps.

[0058] RSRP data originates from the BARTLab Radiomaps dataset, where urban radio propagation is simulated using the Altair Feko electromagnetic modeling tool. Each simulation assumes a building height of 10 meters and includes three 720842A2 antennas installed at 30 meters high, each with a transmit power of 46.00 dBm. Although this dataset supports five frequency bands from 1750 MHz to 5750 MHz, in this embodiment, the investigation is restricted to the 1750 MHz frequency band to ensure spectral consistency across all samples.

[0059] The original dataset contains approximately 2000 simulated RSRP maps with spatial resolutions ranging from (140 x 268) to (561 x 1313) pixels. These maps are constructed on a structured grid aligned with the topography layout extracted from OpenStreetMap. To incorporate environmental context, Sentinel-1 SAR imagery is retrieved from the Google Earth Engine (GEE) platform. Only urban areas with reliable SAR coverage are retained, resulting in a carefully curated multi-modal subset of 1600 samples.

[0060] Prior to training, all input modalities undergo a series of preprocessing steps to ensure spatial alignment and resolution consistency. First, RSRP maps are center-cropped and uniformly resized to (256 x 256) pixels, processed using bilinear interpolation. This standardization facilitates batch processing and architectural consistency. Second, SAR images are temporally averaged across multiple acquisitions to mitigate speckle noise while preserving stable surface scattering patterns. Third, by selecting SAR acquisitions from the same time period as the RSRP simulations, seasonal alignment is enforced, reducing discrepancies due to environmental changes. Finally, building information maps are rasterized and resampled to match the spatial resolution of RSRP and SAR data.

[0061] After preprocessing, each data sample is represented as a spatially aligned triple (Y OBS ,X SAR ,X B ), where Y OBSSparse RSRP measurements, X SAR SAR backscatter intensity map, X B Corresponds to the building information layer. This dataset design enables the proposed CMCTNet to learn the joint spatial relationship between radio frequency propagation and environmental structures, thereby achieving robust reconstruction under high spatial heterogeneity and extreme measurement sparsity.

[0062] In summary, CMCTNet integrates wavelet-based feature extraction, densely connected spatial modeling, and adaptive multi-modal attention to achieve robust and high-resolution REM reconstruction. Its architectural components are specifically tailored to the challenges of cross-modal heterogeneity, signal sparsity, and environmental complexity inherent in real-world wireless sensing scenarios.

[0063] Simulation results evaluation

[0064] In the CMCTNet model, the input data includes three modalities: RSRP data, SAR images, and building height maps, with an input size of 256x256 for each modality. First, the three types of input are independently subjected to preliminary spatial feature extraction through separate wavelet blocks, generating 32-channel low-level feature maps. Next, the RSRP, SAR, and building data are fed into their respective encoder branches, which consist of multiple dense blocks and are downsampled at each layer through max pooling to extract more abstract features. In the RSRP encoder, the features are upsized from 32 channels to 64, 64, 128, and finally to 256-channel deep semantic features. The SAR and building branches also follow the same structural process, extracting their own 256-channel high-level features. At this point, the three modalities have each obtained a deep semantic feature map with a spatial size reduced by half and 256 channels. Subsequently, the high-order features of the three modalities are sent to the fusion attention module, which uses a multi-channel attention mechanism to adaptively fuse complementary information between the three modalities, forming a comprehensive representation feature, and adding a double attention mechanism at the end to enhance the feature selection ability between channels. The fused feature map is then further extracted through two convolution modules to obtain a 512-channel deep fusion representation, and a dropout layer is used for regularization to prevent overfitting. The decoding stage uses a step-by-step upsampling structure: the fusion features are first upsampled to 1 / 8 of the original spatial size (consistent with the last layer before downsampling in the encoder), and then fused with the RSRP and building branch features at the same level. Each step of upsampling and fusion is followed by a dense block to further fuse and recover more spatial details. Specifically, the 512-channel upsampled features are fused with two 256-channel features to form a 1024-channel feature, which is then compressed to 256 channels through a dense block, and the subsequent steps follow the same pattern, finally recovering to 64 channels. In the last stage, the 64-channel decoding features are first reduced to 32 channels through a 7x7 convolution and ReLU activation, and then output through a 3x3 convolution to obtain the final prediction result, i.e., the restored RSRP distribution map, with an output channel of 1, consistent with the input. The entire process achieves high-fidelity reconstruction of the complete electromagnetic environment map from sparse RSRP and auxiliary modalities (SAR and buildings).

[0065] To rigorously evaluate the reconstruction performance of the proposed CMCTNet under sparse sampling conditions, four widely used quantitative indicators are adopted: mean absolute error (MAE), root mean square error (RMSE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). MAE and RMSE are standard indicators of numerical accuracy, while SSIM and PSNR are used to assess the structural similarity and perceptual quality between the reconstructed radio environment map (REM) and the true value. The definitions of these indicators are as follows:

[0066]

[0067]

[0068] These four metrics collectively provide a robust framework for assessing the accuracy and perceptual consistency of the reconstructed REM.

[0069] To improve the generalization ability and robustness of the model under limited training data, data augmentation techniques are applied during the training phase. Each training sample is randomly flipped horizontally, vertically, or rotated by 90 degrees with a 50% probability. These geometric augmentations encourage the model to learn propagation patterns that are invariant to translation and rotation, which is particularly important in heterogeneous urban environments where signal behavior can vary significantly due to layout orientation and structural symmetry.

[0070] To benchmark CMCTNet, it is compared with two representative deep learning-based REM reconstruction methods:

[0071] 1. RobUNet: A UNet-based radio map construction algorithm that utilizes an encoder-decoder architecture to capture multi-scale spatial features. By leveraging hierarchical features, RobUNet achieves better generalization ability under limited measurements in complex urban environments.

[0072] 2. DeepREM: A dual-branch model that combines U-Net and conditional generative adversarial network (CGAN) structures to estimate REM from sparse observations. Notably, DeepREM does not require auxiliary information, making it suitable for scenarios where little environmental prior information is available.

[0073] All comparative experiments are conducted under consistent data partitioning and sampling protocols to ensure fairness.

[0074] Experiments are conducted using uniform random sampling with sampling rates of 5%, 3%, and 1%. For comparison, two baseline methods, DeepREM and RobUNet, are also included. The evaluation focuses on four quantitative metrics—MAE, RMSE, SSIM, and PSNR—that capture the numerical accuracy and perceptual quality of the reconstructed radio environment map.

[0075] Table 1 shows the results for MAE and RMSE. CMCTNet consistently achieves the lowest reconstruction error across all sampling rates. As the sampling rate decreases, the performance gap between CMCTNet and the baselines becomes more pronounced, highlighting the robustness of the proposed model under extreme sparsity conditions. Notably, DeepREM exhibits a significant drop in performance at lower sampling rates due to the lack of structural priors. RobUNet maintains relatively stable performance but still lags behind CMCTNet, especially in low-sampling scenarios.

[0076] Table 1 Comparison of MAE and RMSE under uniform sampling

[0077]

[0078] To further evaluate structural consistency and perceptual fidelity, Table 2 reports the SSIM and PSNR scores at the same sampling rate. CMCTNet once again demonstrates superior performance, producing reconstructions with sharper spatial boundaries and higher visual quality. Even at the extreme 1% sampling rate, CMCTNet maintains an SSIM of 0.876 and a PSNR of 27.39 dB, significantly outperforming RobUNet and DeepREM.

[0079] Table 2 Comparison of SSIM and PSNR under uniform sampling

[0080]

[0081] At the most challenging 1% sampling rate, DeepREM's performance degrades significantly, achieving only 0.498 SSIM and 20.76 dB PSNR. In contrast, CMCTNet continues to produce high-fidelity reconstruction results, demonstrating its strong generalization ability and superior modeling of spatial and structural dependencies.

[0082] Figure 5 A visual comparison of the reconstructed REM at different sampling levels is shown. CMCTNet exhibits a clearer signal transition and more accurate propagation boundaries, especially near building edges.

[0083] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all the implementation methods here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A high-precision radio environment map reconstruction method based on multimodal fusion, characterized in that, include: Step S01: Multiple parallel encoders are used to process multiple input information separately to generate multiple corresponding latent feature representations; Step S 02: The fusion attention module is used to perform cross-modal dynamic fusion of the multiple latent features to obtain a fused latent feature representation, including: constructing a spatial adaptive attention map, and using a dual attention mechanism to simultaneously model spatial and channel dependencies, grouping the spatial adaptive attention map, processing it through spatial or channel attention sub-branches, and finally shuffling and connecting it to promote cross-channel interaction; Step S03: Reconstruct the fused latent feature representation using a decoder, including preserving high-resolution spatial features through skip connections and transferring location information from the encoder to the decoder to obtain a high-precision radio environment map; In step S01, the processing of multiple input information includes: using discrete wavelet transform to perform initial feature extraction on multiple input information, decomposing the input information into multiple frequency bands, and performing multi-level feature extraction processing on the multiple frequency bands through separate convolutional layers to generate latent feature representations; The multiple frequency bands include LL, LH, HL and HH; In step S02, during the cross-modal dynamic fusion, the latent feature representation extracted from the sparse RSRP map is used as the query Q, and the latent feature representation of the fusion extracted from the SAR image and building data is used as the key K' and value V. In step S02, the spatial adaptive attention map is constructed based on the following attention calculation formula: ； In the formula, For RSRP query tensor, and These are the keys and values ​​in modal fusion. It is the dimension of the key vector; Each encoder contains several dense blocks. The forward propagation process within the dense block includes: first, performing a first convolution on the input to obtain feature x1; then concatenating x1 with the original input in the channel dimension as the input to the second convolution block to obtain x2; then concatenating x1, the original input, and x2 in the channel dimension as the input to the third convolution block to obtain the final output x3.

2. The high-precision radio environment map reconstruction method based on multimodal fusion according to claim 1, characterized in that, In step S01, the plurality of parallel encoders includes three parallel encoders.

3. The high-precision radio environment map reconstruction method based on multimodal fusion according to claim 1, characterized in that, In step S01, the multiple input information includes sparse RSRP maps, SAR images, and building data.

4. The high-precision radio environment map reconstruction method based on multimodal fusion according to claim 1, characterized in that, In step S03, the decoder outputs a dense prediction map, which is used to estimate the complete RSRP distribution within the discrete region; The dense prediction map is as follows: in, The complete RSRP distribution is estimated using a multimodal deep learning framework, N. x and N y These represent the number of grid cells along the x and y directions after the target region is discretized, respectively.

Images