Abstract
With the increasing demand for real-time channel state information (CSI) in sixth-generation (6G) mobile communication networks, channel knowledge map (CKM) emerges as a promising technique, offering a site-specific database that enables environment-awareness and significantly enhances communication and sensing performance by leveraging a priori wireless channel knowledge. However, efficient construction and utilization of CKMs require high-quality, massive, and location-specific channel knowledge data that accurately reflects the real-world environments. Inspired by the great success of ImageNet dataset in advancing computer vision and image understanding in artificial intelligence (AI) community, we introduce CKMImageNet, a dataset developed to bridge AI and environment-aware wireless communications and sensing by integrating location-specific channel knowledge data, high-fidelity environmental maps, and their visual representations. CKMImageNet supports a wide range of AI-driven approaches for CKM construction with spatially consistent and location-specific channel knowledge data, including both supervised and unsupervised, as well as discriminative and generative AI methods.
Introduction
In the upcoming era of 6G, mobile networks will be denser and more complex than ever. Traditional methods for acquiring Channel State Information (CSI)—the data describing how a signal travels from transmitter to receiver—are becoming too slow and "expensive" in terms of overhead.
The solution is a Channel Knowledge Map (CKM): a site-specific database that stores a priori wireless knowledge. However, building an accurate CKM requires massive amounts of high-quality, location-tagged data. Existing datasets are often limited; they might only provide signal strength (path loss) without directional data, or they lack the visual environmental context (like building layouts) that AI needs to learn why a signal behaves a certain way.
As shown in the comparison above, CKMImageNet aims to be the "ImageNet of the wireless world," providing the scale and diversity needed to train sophisticated AI models for 6G by offering a "full-stack" of information (location + signal + angles + delays + visual maps) that its predecessors lack.
The CKMImageNet Dataset
CKMImageNet is a massive, high-fidelity dataset designed to revolutionize 6G wireless research by bridging the gap between Artificial Intelligence and electromagnetic propagation. By providing location-specific channel data paired with visual environmental maps, it enables the development of "environment-aware" communication systems that can predict signal behavior based on the physical world around them.
The dataset features over 10,000 environment maps, each spanning 128m × 128m. It includes multi-dimensional data: channel gain, Angle of Arrival (AoA), Angle of Departure (AoD), and propagation delay. Each study area in the dataset produces up to 1,971,216 data points (16 base stations × 123,201 receiver locations). The hierarchical structure allows for diverse scenarios, from urban centers like London and Nanjing to rural suburbs and industrial factory floors.
Methodology
The authors used a multi-step pipeline to generate high-fidelity synthetic data that mimics real-world physics:
- Environment Modeling: They used OpenStreetMap (OSM) to select real-world locations and processed them in Blender to create detailed 3D models of buildings and terrain.
- Ray-Tracing Simulations: Using professional tools like Wireless Insite and the open-source Sionna framework, they simulated how radio waves at 28GHz interact with these 3D environments, accounting for reflections, diffractions, and scattering.
- Data Transformation: To make this data "AI-friendly," they converted numerical channel parameters into normalized grayscale images.
- Multi-Resolution Scaling: The dataset provides images in three sizes (32×32, 64×64, and 128×128 pixels) to support different application needs, from resource-constrained edge devices to high-precision network planning.
Key Insights and Results
- The "Visual CKM" Concept: By representing complex electromagnetic data as grayscale images, researchers can leverage powerful, pre-trained Computer Vision architectures (like CNNs, GANs, and Diffusion Models) that were originally designed for natural images.
- Environment-Data Fusion: The dataset couples signal values with binary obstacle matrices (building layouts). This allows AI to learn the causal relationship between physical objects and signal "shadows" or multipath reflections.
- Spatial Consistency: Unlike statistical models that rely on randomness, CKMImageNet is built on ray-tracing, ensuring that the data is spatially consistent—meaning the signal changes logically as a user moves through the environment.
The authors demonstrated that CKMImageNet can be used for diverse tasks, including CKM Inpainting (filling missing data), Super-Resolution (upscaling channel maps), and Enhanced Estimation for XL-MIMO systems.
Conclusion and Future Work
While the current dataset is primarily synthetic, the authors have designed a Data Sharing Platform to allow researchers worldwide to contribute location-tagged measurement data. Future iterations of CKMImageNet plan to include frequency diversity (expanding beyond 28GHz), dynamic environments (weather, time-of-day), and multi-level infrastructure accounting for varying heights of transmitters and receivers.
Summary updated