💻 computer science

Environment-specific, data-driven radiation source detection, localization and tracking via a sensor network and ensembles of randomized decision trees

This paper presents a data-driven approach using sensor networks and ensembles of randomized decision trees to accurately detect, localize, and track radioactive sources in complex environments, demonstrating superior performance over traditional methods while eliminating the need for detailed radiation transport modeling.

Original authors: Andrew White, Cory Simon, Haori Yang

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew White, Cory Simon, Haori Yang

Original paper licensed under CC BY 4.0 (https://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

Imagine you are trying to find a lost, glowing marble hidden inside a room full of furniture, walls, and boxes. The marble emits a faint, invisible "hum" (radiation) that gets blocked or bounced around by the furniture. If you just stand in the middle of the room and listen, you might hear the hum, but you won't know exactly where it is because the furniture distorts the sound.

This paper describes a clever way to solve that problem using a team of "ears" (sensors) and a smart "brain" (computer algorithm) that learns by experience rather than by doing complex physics math.

Here is the breakdown of their approach:

1. The Setup: A Room Full of Obstacles

The researchers built a small-scale model of a messy environment on a table. They placed obstacles made of different materials—like lead (which blocks radiation), wood, and cardboard—to act like walls and furniture. These obstacles absorb or scatter the radiation, making it hard to pinpoint the source using simple rules.

They set up eight "ears" (radiation sensors) around the room. These sensors listen for the "hum" of a radioactive source (specifically a small amount of Cesium-137).

2. The Problem: Why Old Methods Fail

Traditionally, finding a source is like trying to find a sound by guessing where it's loudest.

The "Weighted Centroid" method: This is like assuming the source is just the average location of the three loudest ears. It's a simple guess that often fails when furniture blocks the sound.
The "Multilateration" method: This tries to calculate the distance based on how loud the sound is, assuming the air is empty. But in a room full of furniture, the sound gets muffled, so this math gets confused and leads to wrong answers.

3. The Solution: Teaching a Computer to "See" the Pattern

Instead of trying to write a complex math equation to account for every piece of furniture, the researchers decided to teach a computer by showing it examples.

The Training Phase: They placed the glowing marble in 100 different spots around the table. Every time they placed it, they recorded what all eight sensors heard. They also recorded what the sensors heard when the marble was not there (background noise).
The "Brain": They used a type of computer learning called Randomized Decision Trees. Imagine a team of 100 detectives, each asking a simple question like, "Did Sensor #4 hear more than 50 beeps?" or "Is Sensor #1 quieter than Sensor #2?"
- Each detective makes a guess based on their own set of questions.
- The final answer is the average of all the detectives' guesses.
- Because the detectives are "randomized," they don't all make the same mistakes. This makes the group very good at spotting patterns that a single detective (or a simple math formula) would miss.

4. How It Works in Real-Time

Once the computer is trained, it acts like a two-step security guard:

Step 1 (The Detector): The computer looks at the sensors' current readings. It asks, "Is the glowing marble here, or is it just background noise?" It is very good at this, rarely raising a false alarm.
Step 2 (The Tracker): If it says "Yes, the marble is here," it immediately switches to the second mode. It looks at the pattern of the sensors and says, "Based on what we learned, the marble is likely at this specific spot."

5. The Results: Smarter Than the Math

When they tested their system:

Accuracy: The computer found the source much more accurately than the traditional math methods. While the old methods got lost in the "furniture," the computer learned exactly how the furniture changed the sound.
Confidence: The computer didn't just give a guess; it gave a "confidence circle." If the sensors were confused, the circle got bigger, telling the user, "I think it's around here, but I'm not 100% sure."
Tracking: In a final test, they moved the source slowly across the table. The system successfully traced the path of the moving source, like a GPS tracking a slow-moving car.
Sensor Importance: The system even told them which sensors were the most important. It found that some sensors were redundant (doing the same job as others), suggesting you could remove them to save money without losing much accuracy.

Summary

In short, instead of trying to calculate the physics of how radiation bounces off every piece of furniture, the researchers built a "smart team" that learned by watching thousands of examples. This allowed them to find a hidden radioactive source in a cluttered room with high accuracy, outperforming traditional math-based methods that get confused by obstacles.

What the paper doesn't claim:

It does not claim this works for any radioactive source; it was trained specifically on one type (Cesium-137) with a specific strength.
It does not claim to handle multiple sources at once.
It does not claim to work in a massive city; it was tested on a tabletop model.
It does not claim to work if the source moves very fast (the sensors need 60 seconds to listen, so fast movement would blur the signal).

Technical Summary: Environment-Specific, Data-Driven Radiation Source Detection, Localization, and Tracking

Problem Statement
Radioactive sources are essential for medical, industrial, and agricultural applications but pose significant safety and security risks if abandoned, misplaced, or stolen. Continuous surveillance is required to detect, localize, and track these sources in complex environments. Traditional physics-based localization methods often struggle in cluttered settings where obstacles (e.g., walls, machinery) absorb and scatter radiation, complicating the interpretation of sensor data. Furthermore, creating accurate, computationally efficient models of radiation transport in such complex, obstacle-filled environments is difficult. The authors address the challenge of detecting and localizing a specific radioactive source in a complex, obstacle-rich environment without relying on detailed physics-based transport simulations.

Methodology
The authors propose a data-driven framework utilizing a spatially distributed network of stationary radiation sensors and supervised machine learning.

Experimental Setup: A tabletop environment (42 in × 33 in) was constructed with obstacles made of lead, cardboard, fiberboard, and wood to simulate a complex, cluttered space. A network of eight identical NaI(Tl) scintillation sensors was distributed across the environment.
Data Collection:
- Source Exposure: 100 experiments were conducted placing a 7.2 µCi $^{137}$ Cs source at diverse locations (selected via Latin Hypercube sampling and targeted error reduction) to generate positive training examples.
- Background: 61 experiments were conducted without the source to generate negative training examples.
- Tracking: A separate experiment tracked a slow-moving source along a path of 22 discrete steps to test dynamic localization.
Machine Learning Models: Two ensembles of Extremely Randomized Trees (ERTs) were employed:
1. Detector (Classifier): Trained on both background and source-exposure data to predict the binary presence or absence of the source.
2. Localizer (Regressor): Trained exclusively on source-exposure data to predict the $(x, y)$ coordinates of the source.
- The ERT approach was chosen for its resistance to overfitting, speed, built-in feature importance scoring, and ability to provide empirical uncertainty estimates via prediction variance among the tree ensemble.
Evaluation: Performance was assessed using Leave-One-Out Cross-Validation (LOOCV). The system was benchmarked against three baselines: (1) Weighted Centroid, (2) Multilateration (using an empirically fitted distance-response curve), and (3) Random Guessing.

Key Results

Detection Performance: The ERT classifier achieved an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.99 and an Area Under the Precision-Recall Curve (AUC-PR) of 0.995. It produced zero false positives. Five false negatives occurred, all in the bottom-right corner where lead blocks heavily shielded the sensors from the source.
Localization Performance: The ERT regressor achieved a mean localization error of 3.4 inches. This significantly outperformed the multilateration baseline (4.9 inches), the weighted centroid method (14.9 inches), and random guessing (20.3 inches).
Learning Curve: The study determined that approximately 100 training examples were sufficient to reach a performance plateau. Notably, more than 50 training examples were required for the ERT ensemble to outperform the physics-based multilateration method.
Uncertainty and Sensor Importance: The ensemble provided quantified uncertainty (via prediction variance) which correlated positively with actual LOOCV errors. Sensor importance scores identified sensors 1, 2, and 4 as most critical for localization, while sensors 3, 5, and 7 were deemed less critical, offering insights for network optimization.
Tracking: The system successfully reconstructed the path of a slow-moving source in a capstone experiment, with errors increasing primarily when the source was near the environment's periphery.

Significance and Claims
The paper claims to demonstrate a viable, environment-tailored approach for radiation source surveillance that circumvents the need for computationally expensive and detailed radiation transport models. By training on experimental data, the system implicitly learns the effects of radiation absorption and scattering caused by environmental obstacles.

The authors emphasize that while the system is currently tailored to a single source type ( $^{137}$ Cs) and a specific layout, the data-driven framework offers a practical alternative for complex environments where physics-based modeling is prohibitive. The integration of machine learning allows for accurate localization with quantified uncertainty and provides actionable metrics (sensor importance) for optimizing sensor network configurations.

Limitations and Future Directions
The authors modestly acknowledge several limitations:

The system is currently specific to one source composition and activity; different sources would require new training data.
It cannot localize multiple simultaneous sources.
Changes in the environment layout or sensor positions necessitate retraining.
Scalability to very large areas is limited by the need for high sensor density to ensure coverage.

Future work proposed by the authors includes adapting the system to varying source characteristics, handling multiple sources, integrating contextual data (e.g., video), and fusing data-driven methods with physics-based simulations to reduce training data requirements.