🔭 astrophysics

Tropospheric Ducting Prediction based on GFS Model Data at Inyarrimanha Ilgari Bundara, the CSIRO Murchison Radio-astronomy Observatory

This paper presents and validates a tropospheric ducting prediction method for the CSIRO Murchison Radio-astronomy Observatory using GFS model data and machine learning, which successfully forecasts RFI events to enable adaptive observation scheduling.

Original authors: Balthasar Indermuehle, Hajime Suzuki

Published 2026-08-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Balthasar Indermuehle, Hajime Suzuki

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Highways of the Sky

Imagine the sky above us not as an empty void, but as a giant, invisible ocean of air. Usually, this ocean is calm, letting radio waves travel in straight lines until they hit the horizon and stop, just like a flashlight beam fading into the distance. But sometimes, the weather plays a trick. When a layer of warm air sits on top of cooler air near the ground, it acts like a giant, invisible tunnel or a "duct." This is called tropospheric ducting.

Think of it like a wave in a swimming pool. If the water is flat, the wave spreads out and dies. But if you build a wall around the pool, the wave gets trapped and can travel all the way to the other side without losing much energy. In the atmosphere, these "walls" are made of temperature and humidity. When they form, they can trap radio signals—like TV broadcasts, cell phone chatter, or ship radios—and guide them hundreds of kilometers over the horizon.

This matters because some places on Earth are built to be "radio quiet zones," like a library for telescopes that listen to the faint whispers of the universe. If these invisible sky-tunnels open up, they can smuggle noisy radio signals from distant cities right into the telescope's sensitive ears, drowning out the cosmic secrets scientists are trying to hear. The big question is: Can we predict when these sky-tunnels will open, so the telescope can close its ears just in time?

The Paper's Story: Predicting the Sky's Secret Tunnels

This paper is a detective story about the CSIRO Murchison Radio-astronomy Observatory in Western Australia, a place so quiet it's one of the best spots on Earth to listen to the universe. The scientists, Balthasar Indermuehle and Hajime Suzuki, wanted to solve a mystery: Why does the telescope sometimes hear strange radio noise from hundreds of kilometers away, even though there are no transmitters nearby? They suspected the culprit was tropospheric ducting, and they set out to build a crystal ball to predict it.

The Detective Work: Reading the Weather Map
The team didn't just guess; they used a massive weather forecast model called the GFS (Global Forecast System), which is like a super-computer that simulates the atmosphere's temperature and humidity. They looked at the data to find the specific "recipe" for a sky-tunnel: a sharp drop in how much the air bends radio waves.

They built a prediction system that checks the weather every hour. If the computer sees the right conditions—like a temperature inversion that acts as a lid—it flags a potential "ducting event." They tested this system against seven years of real data, watching 25 different radio channels ranging from 88 MHz (FM radio) up to 2.68 GHz (mobile phone speeds).

The Big Findings: A Good Crystal Ball
The results were surprisingly good. The system could predict when interference would happen with a skill score of 0.724 (on a scale where 0.5 is a coin flip and 1.0 is perfect). This means the system is far better than guessing.

It works for days: The prediction stays accurate for up to three days in advance. This is huge because it means astronomers can plan their observations ahead of time, switching to "safe" frequencies when a sky-tunnel is forecasted to open.
It works for different types of signals: The system was great at predicting interference for mobile phone bands and TV signals. However, it found that FM radio channels (like 88–97 MHz) were much harder to predict. The paper suggests these signals aren't coming from far away via sky-tunnels, but are likely just weak, local signals that are barely strong enough to be heard.
The "Null" Result: Interestingly, one channel at 460 MHz showed no skill at all. The paper calls this a "well-powered null," meaning the model's predictions were no better than random chance; it could not distinguish between a real event and a false alarm on this specific frequency. This proves the model is honest enough to admit when it doesn't know, rather than forcing a prediction where none is possible.

Proving It's Real: The "Ground Truth"
To make sure they weren't just seeing ghosts in the data, the team used two clever tricks to prove the sky-tunnels were real:

The Cell Phone Scanner: They used a device that decodes mobile phone signals. When the system predicted a duct, the scanner actually found cell towers from hundreds of kilometers away (up to 589 km) that were suddenly visible. The phone networks matched the predictions perfectly.
The Ship Tracker: They tracked ships at sea using their Automatic Identification System (AIS). Ships are like floating radio beacons. The team found that when the system predicted a duct, the ships' signals were much stronger and traveled further over land than usual. Specifically, the "land-segment duct fraction" was 67% higher during these events compared to normal times.

What the Paper Rules Out
The authors were careful to rule out other possibilities. They checked if the noise came from aircraft scattering signals or from sporadic-E (a weird ionosphere effect), but the data didn't fit. They also proved that simply having a higher-resolution weather map didn't necessarily make the prediction better; sometimes, the standard map was actually more reliable because it matched the real-world data better.

The Bottom Line
This paper doesn't just say "it happens"; it gives a working tool. The team has already turned their method into a live forecast service for the observatory. Now, instead of being surprised by radio noise, the astronomers can look at a forecast, see a "ducting alert," and move their observations to a quiet frequency. It's like having a weather app that tells you when the sky is going to open a secret highway for radio waves, allowing the telescope to keep its ears open for the universe's quietest whispers.

Technical Summary: Tropospheric Ducting Prediction at Inyarrimanha Ilgari Bundara

Problem Statement
The CSIRO Murchison Radio-astronomy Observatory (Inyarrimanha Ilgari Bundara) in Western Australia is situated within the Australian Radio Quiet Zone (ARQZWA), designed to shield sensitive instruments like ASKAP, MWA, and SKA-Low from terrestrial radio-frequency interference (RFI). Despite its remote location (300–600 km from major population centers) and regulatory protections, the site experiences episodic tropospheric ducting. This phenomenon, driven by temperature inversions and humidity lapses, traps VHF and UHF radio waves, guiding them hundreds of kilometers beyond the radio horizon. These ducted signals, originating from licensed transmitters in Perth and Geraldton, elevate the noise floor, saturate receiver chains, and contaminate scientific data. Currently, no published system predicts ducting-borne interference specifically for observatory operations, leaving astronomers unable to adaptively schedule observations to avoid affected frequencies.

Methodology
The authors developed a prediction pipeline that couples numerical weather prediction (NWP) data with ITU-R propagation models to estimate interference levels at the observatory.

Atmospheric Inputs: The system utilizes vertical profiles from the NOAA Global Forecast System (GFS) at 0.25° resolution. From these profiles, radio refractivity ( $N$ ) and modified refractivity ( $M$ ) are calculated to derive duct diagnostics, including the minimum vertical gradient of $M$ ($dM/dz$) in the lowest 3 km, duct base/top heights, and integrated duct depth.
Source Attribution: Candidate interference sources are identified using the ACMA Register of Radiocommunications Licences (RRL). Transmitters are matched to monitored channels based on emission bandwidth overlap.
Propagation Modeling: Basic transmission loss ( $L_b$ $L_{b}$ ) is computed for every transmitter-to-site path using:
- ITU-R P.452-16 for frequencies $\ge$ 100 MHz.
- ITU-R P.1812-6 for frequencies < 100 MHz.
- These models are driven by GFS-derived refractivity lapse rates ( $\Delta N$ ) along the path over Copernicus GLO-90 terrain.
- The models are treated as statistical tools where the time percentage parameter ( $p$ ) acts as a regime selector for anomalous propagation rather than a strict probability.
Validation Data: The system is validated against seven years of continuous spectrum monitoring (2019–2026) across 25 channels (88 MHz to 2.68 GHz). An "event" is defined as simultaneous elevation in at least three channels to suppress local artifacts.
Machine Learning Augmentation: A pooled gradient-boosted model (GBM) was trained on the same physical inputs (analytical predictions, raw duct diagnostics, and climatology harmonics) to assess if non-linear coupling could improve skill.

Key Contributions and Innovations

First Operational Ducting Forecast: This work presents the first system to connect NWP duct diagnostics directly to measured interference at a radio-quiet observatory, enabling a live forecast service.
Bandwidth Discrimination for Self-Interference: The authors successfully recovered two 2.6 GHz channels (2650 and 2680 MHz) that were previously masked by the observatory's own ASKAP leakage. By discriminating based on signal bandwidth (broadband licensed pedestal vs. narrowband leakage) rather than amplitude, they isolated the ducted signals.
Stratified Skill Scoring: To prevent the model's skill from being inflated by shared seasonal or diurnal cycles, the authors employ a stratified Area Under the Curve (AUC) metric, comparing event hours only against non-event hours within the same month and three-hour block.
Duct-Corridor Connectivity: A novel metric was introduced to measure the continuity of duct layers along a path. While standalone performance was weak, augmenting the GBM with this feature improved test AUC, particularly for land-based paths where GFS under-detects ducts compared to ERA5.
Cross-Model Validation: The analytical predictor was validated against a different GFS model version (v15 vs. v16) and different atmospheric models (ERA5, BARRA-C2), demonstrating robustness to model version changes and highlighting the importance of horizontal resolution over vertical resolution for this application.

Results

Analytical Performance: Across 25 channels, the median stratified AUC is 0.724, with 24 of 25 confidence intervals excluding chance. The strongest performance is observed in mobile and IMT bands (e.g., 2680 MHz: 0.779; 2125 MHz: 0.761).
Forecast Lead Time: Skill remains flat (approx. 0.74) through three days of forecast lead time, then degrades gradually.
Machine Learning Gains: The pooled GBM, augmented by the corridor connectivity metric, achieved a mean test AUC of 0.806 across all 25 channels and 0.799 on the 16 channels where the analytical predictor was already skilled.
Ground Truth Verification:
- LTE Decoding: 86–97% of over-the-horizon LTE cell detections fell within flagged event hours, with decoded network identities matching licensing records.
- AIS Ship Receptions: At 162 MHz, ship receptions showed a 67% higher land-segment duct fraction compared to control samples, confirming the GFS diagnostics capture VHF ducting.
Channel Specifics:
- 460 MHz: A "well-powered null" with an AUC indistinguishable from chance (0.536).
- FM Channels (88–97 MHz): Showed weak but non-zero skill (AUC ~0.60–0.65). The authors attribute this to near-threshold reception of local low-power sources modulated by ordinary refractive variability, rather than strong ducting.
- 2.6 GHz: Successfully recovered from self-interference, showing strong seasonal cycles consistent with ducting.

Significance and Claims
The paper claims to provide a practical, site-agnostic recipe for predicting ducting-borne interference using open data (GFS, ACMA RRL, Copernicus terrain) and open-source propagation implementations. The primary significance is operational: the method enables adaptive scheduling of observations away from frequencies affected by forecast ducting and drives a live forecast service at the observatory.

The authors are modest regarding probabilistic calibration; while the system discriminates well (high AUC), several channels (e.g., 2680 MHz) do not yield usable probabilistic information (negative Brier skill) without specific re-fitting. The study emphasizes that the physical claims rest on the deterministic ITU-R layer, while the machine learning layer serves to extract additional skill from the same inputs. The work demonstrates that ducting is a predictable, dominant source of interference in the region, and that its prediction can be integrated into observatory operations to protect data quality.

The Invisible Highways of the Sky

The Paper's Story: Predicting the Sky's Secret Tunnels

More like this