[ Research ] · Machine learning
Indoor 5G coverage from drone data using graph attention
Our ACM Web Conference 2026 companion paper infers indoor 5G signal quality from drone readings taken outside, using a dynamic graph attention network.
On this page
Indoor 5G coverage is hard to predict, because the structure of tall buildings weakens radio signals in ways that simple models miss. Our paper “Spatial-Temporal Inference of Indoor RSSI and CQI for 3D Coverage Mapping using Dynamic Graph Attention Networks” was led by David Hason Rudd and written with Johan Sebastian Ramirez Vallejo, Md. Rafiqul Islam, Andrew Levula and Cesar Sanin. It appears in the Companion Proceedings of the ACM Web Conference 2026 (ACM, pp. 251-257). The paper treats unreliable indoor 5G as a significant weakness in modern urban infrastructure and proposes a way to estimate signal quality inside a building from drone measurements taken outside it.
How the dynamic graph attention model works
Drones, or unmanned aerial vehicles, are good at measuring radio frequency signals around the outside of a building. Working out what is happening inside from a sparse set of outside readings is much harder. It is an inverse problem: the model has to reason backwards from limited evidence to conditions it cannot observe directly.
Our approach uses a Dynamic Graph Attention Network (DGAT). A graph neural network represents data as nodes joined by edges and passes information along those edges. Attention lets the model give more weight to some neighbours than to others.
The central design choice is how the graph is built. Conventional models connect measurement points by Euclidean, or straight-line, distance, on the assumption that points close together will have similar signals. Inside a complex building, signal propagation is non-linear, so physical closeness is a weak guide. Our model instead learns a hidden connectivity structure from feature affinity. It links points whose measured characteristics are alike rather than points that simply sit near each other.
The learned graph then feeds three further components:
- a deep residual propagation encoder, which passes information through the learned structure
- a dual-branch decoder, with one branch for RSSI (Received Signal Strength Indicator) and one for CQI (Channel Quality Indicator), because these measures of signal strength and channel quality behave differently in physical terms
- a meta-ensemble fusion layer, which combines outputs to reduce the variance of the predictions
What the results on a multi-level 5G dataset show
We tested the framework on a multi-level 5G dataset and compared it with existing location-aware baselines. It reduced root mean square error (RMSE), a standard measure of how far predictions fall from measured values, by 76% for RSSI and 89% for CQI. Its R² scores, which show how much of the variation in the measured values the model accounts for, were above 0.99.
Put simply, the predictions tracked the measured indoor values very closely on this dataset. The relative improvement over location-aware methods was largest for channel quality.
Why learned connectivity matters for indoor 5G planning
The paper extends an earlier study, led by David Hason Rudd with several of the same authors and published in Procedia Computer Science in 2025. That study sets out the practical problem. Higher-frequency 5G bands lose much of their strength passing through building materials, and modern buildings have complex layouts. Checking in-building coverage against mobile network operator standards has traditionally meant indoor walk tests, which are expensive and slow.
The earlier work used drones carrying cellular network scanners to record RSSI and CQI from the main 5G macrocell outside a building. Its Location-Aware Deep Neural Network (LA-DNN) combined those readings with spatial coordinates, free-space path loss models and material-specific attenuation factors. The model was validated against indoor walk-test data and produced 3D heatmaps showing coverage gaps and RF null zones floor by floor. The aim was to reduce indoor measurement during initial assessments, before In-Building Coverage systems are deployed.
The new paper shifts the emphasis from location to learned relationships. Instead of assuming that nearby points should inform each other, the DGAT lets the data decide which points are related. We present the framework as a link between non-invasive aerial sensing and indoor coverage estimation, and as a foundation for next-generation urban wireless planning. If the accuracy holds in other buildings, planners could need fewer site visits when deciding where in-building systems are needed.
My perspective: graph learning, automation and teaching
I think this is a good piece of work, and it connects with three areas I work in.
The first is how relationships are represented. In our cybersecurity ontology paper, the relationships between concepts are made explicit so that people and systems can share them. Here, the relationships between measurement points are learned from data. Both treat the choice of connections as central to the result. That raises a question: when should structure be specified by people, and when should it be learned?
The second is automation. Replacing part of a slow manual survey with drone sensing and a learned model is one example of a wider shift towards automating routine measurement work. This raises a practical question for industry: how far can predictions like these stand in for direct measurement, and who checks them before they inform decisions about where to invest in infrastructure?
The third is teaching. For students in information technology programs, the paper is a useful case study in why the way data is structured matters as much as the choice of model. A problem that looks spatial at first turns out to be better described by learned similarity than by distance. This raises a question for course design: should networking and machine learning units share more examples like this, rather than being taught separately?
Limits of this summary
This post draws on the published abstract and bibliographic records, not the full paper, so it leaves out some important detail:
- The abstract does not describe the dataset in depth. It does not give the building, the number of levels, the number of measurements or how the data was split for testing.
- It does not name the location-aware baselines. They should not be assumed to be the earlier LA-DNN, and the error figures from the two papers should not be compared directly.
- The title refers to spatial-temporal inference, but the abstract does not explain the temporal side of the model, so it is not covered here.
- Results from one multi-level dataset may not carry over to buildings with different materials, layouts or network settings.
The full method and evaluation are in the paper itself, available through the ACM Digital Library.
Sources and further reading
- Spatial-Temporal Inference of Indoor RSSI and CQI for 3D Coverage Mapping using Dynamic Graph Attention Networks, David Hason Rudd, A B M Mehedi Hasan, Johan Sebastian Ramirez Vallejo, Md. Rafiqul Islam, Andrew Levula and Cesar Sanin. Companion Proceedings of the ACM Web Conference 2026 (WWW '26), ACM, 2026, pp. 251-257
- Location-Aware Deep Neural Network for Predicting Indoor 5G RSSI and CQI Using Drone-Based External RF Sensing, David Hason Rudd, Cesar Sanin, Koh Ming En, Xingyi Gao, Md Rafiqul Islam, Mehedi Hasan, Xianzhi Wang, Angela Huo and Guandong Xu. Procedia Computer Science, vol. 270, Elsevier, 2025, pp. 4765-4775