RadioFormer: A Multiple-Granularity Radio Map Estimation Transformer with 1\textpertenthousand Spatial Sampling

TL;DR

RadioFormer tackles radio map estimation under extremely sparse sampling by deploying a multiple-granularity transformer that separately encodes pixel-level observations and patch-level building geometries, then fuses them with cross-attention to produce dense radio maps. The method leverages a dual-stream self-attention (DSA) to extract multi-granular features and a cross-stream cross-attention (CCA) module to integrate pixel and building information, yielding accurate maps with lower computational cost than pixel-wise approaches. Evaluations on the RadioMapSeer dataset show state-of-the-art performance across various sampling strategies, with strong generalization and zero-shot capabilities, underscoring practical applicability in resource-constrained deployments. Overall, RadioFormer provides an efficient, scalable framework for spectrum situation generation that can operate effectively with minimal observation nodes and complex urban environments, aided by a public code release.

Abstract

The task of radio map estimation aims to generate a dense representation of electromagnetic spectrum quantities, such as the received signal strength at each grid point within a geographic region, based on measurements from a subset of spatially distributed nodes (represented as pixels). Recently, deep vision models such as the U-Net have been adapted to radio map estimation, whose effectiveness can be guaranteed with sufficient spatial observations (typically 0.01% to 1% of pixels) in each map, to model local dependency of observed signal power. However, such a setting of sufficient measurements can be less practical in real-world scenarios, where extreme sparsity in spatial sampling can be widely encountered. To address this challenge, we propose RadioFormer, a novel multiple-granularity transformer designed to handle the constraints posed by spatial sparse observations. Our RadioFormer, through a dual-stream self-attention (DSA) module, can respectively discover the correlation of pixel-wise observed signal power and also learn patch-wise buildings' geometries in a style of multiple granularities, which are integrated into multi-scale representations of radio maps by a cross stream cross-attention (CCA) module. Extensive experiments on the public RadioMapSeer dataset demonstrate that RadioFormer outperforms state-of-the-art methods in radio map estimation while maintaining the lowest computational cost. Furthermore, the proposed approach exhibits exceptional generalization capabilities and robust zero-shot performance, underscoring its potential to advance radio map estimation in a more practical setting with very limited observation nodes.

Figures (7)

• Figure 1: Several methods have been proposed for generating common spectrum scenarios: a. Physical Method. This approach leverages physical models or prior knowledge of physical principles to generate a radio map, using building maps and transmitter information as inputs. b. Vision Model. In this method, a sampling map is combined with a building map, and a vision model is then used to predict the distribution of spectrum scenarios. c. Pixel Transformer. This technique generates the spectrum distribution by inferring the values of unknown points based on the known data points and processing them iteratively. d. RadioFormer. A novel approach introduced by our work enhances prediction by incorporating building map information to refine the interactions between known points, resulting in a comprehensive radio map.
• Figure 2: The workflow for the RadioFormer: Initially, observation points' position and value information are represented as vectors. These vectors are aggregated to derive the features of the sampling points. Subsequently, self-attention mechanisms facilitate the interaction and fusion of information among the sampling points. This is followed by integrating these fused features with the building map features, which have been encoded using a ViT encoder. This fusion process results in the final prediction map for the situational distribution.
• Figure 3: Visualization results of various models under 3 different sampling categories. The first column represents three sampling methods. We use yellow dots in these visualizations to indicate the positions of the observation points and red lines to partition the building into distinct regions, thereby highlighting the differences between the three sampling strategies.
• Figure 4: Attention map visualization: We visualized the attention map of feature fusion by presenting two images side by side. The image on the left displays the topographic map, with the positions of the target observation points marked by blue five-pointed stars, while green points mark other observation points. The image on the right illustrates the attention distribution of the observation points for the building map features. We provide 2 examples with different sampling categories. This visualization provides insight into how the model directs its focus when integrating information from different sources.
• Figure 5: The relationship between model performance and the number of observation points during training. Sub-figures 1, 2, and 3 present the RMSE $\downarrow$, SSIM $\uparrow$, and PSNR $\uparrow$ results. In each sub-figure, we have delineated two regions using gray dashed and dotted lines to compare the performance of RadioFormer and other models.