📊 statistics

A Probabilistic Framework for Reconstructing Sparse UAV-Based Radiation Monitoring Data Using Gaussian Process Regression and Uncertainty Quantification

This study proposes an uncertainty-aware probabilistic framework that utilizes Gaussian Process Regression to effectively reconstruct sparse UAV-based radiation monitoring data from the Chornobyl Exclusion Zone, outperforming traditional interpolation methods in accuracy while providing essential spatial uncertainty estimates for informed environmental decision-making.

Original authors: Andrii Bondarchuk, Tetiana Nosenko, Yurii Zabulonov

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Andrii Bondarchuk, Tetiana Nosenko, Yurii Zabulonov

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 paint a detailed map of a forest, but you can only walk a few specific paths and take a photo of the ground every few steps. You have a lot of gaps between your photos. If you just connect the dots with straight lines, you might miss a hidden patch of poisonous mushrooms or a clear spring. This is exactly the problem the researchers faced with radiation monitoring in the Chernobyl Exclusion Zone.

Here is a simple breakdown of what they did, using everyday analogies:

The Problem: The "Patchy" Map

The team used drones (UAVs) equipped with sensors to fly over a radioactive area and measure radiation levels. However, the drones couldn't fly everywhere at once. They followed specific routes, leaving large empty spaces between them.

The Challenge: They had a "sparse" dataset—like a puzzle with most of the pieces missing. They needed to guess what the radiation levels were in the empty spaces without just guessing blindly.
The Risk: If they guessed wrong, they might miss a dangerous "hotspot" or waste time checking a safe area.

The Solution: Two Different Ways to Fill the Gaps

The researchers tested two different methods to fill in the missing parts of the map.

1. The "Neighborly" Guess (Ordinary Kriging)
Think of this like a game of "Hot and Cold." If you know the temperature at your house and your neighbor's house, you can make a pretty good guess about the temperature on the street between you.

How it works: This method looks at the distance between measurement points. It assumes that points close together are more similar than points far apart. It draws a smooth curve connecting the known dots.
The Result: It worked well, but it was a bit like a standard GPS map: it gave you a route, but it didn't tell you how confident the GPS was that the route was correct.

2. The "Smart Detective" (Gaussian Process Regression)
This method is like a detective who not only solves the case but also writes down a "confidence score" for every clue.

How it works: Instead of just drawing a line, this mathematical model learns the pattern of the radiation. It asks, "Based on the shape of the data I've seen, how likely is it that this empty space has high radiation?"
The Superpower: It doesn't just give you a number; it gives you a confidence level. It can say, "I am 99% sure this area is safe," or "I am only 50% sure about this spot because I haven't flown a drone there yet."

The Process: Cleaning the Data First

Before trying to fill in the map, the researchers had to clean the data.

The Noise: Drones flying through the air pick up "static" or random glitches (like a radio station with static noise).
The Filter: They used a "median filter," which is like a bouncer at a club. If one data point is wildly different from everyone else (an outlier), the bouncer kicks it out so it doesn't ruin the party. They also averaged the data into a grid, making the messy, uneven points into a neat, organized table.

The Results: Who Won?

They tested both methods against the actual data they had collected.

The Winner: The "Smart Detective" (Gaussian Process Regression) was the best. It made fewer mistakes (lower error rates) than the "Neighborly" method and the basic "connect-the-dots" method.
The Bonus: The Smart Detective provided a second map showing uncertainty.
- Low Uncertainty (Dark Blue): Areas where the drone flew close by. The model is very confident here.
- High Uncertainty (Red): Areas far from the flight paths. The model is saying, "I'm guessing here, so be careful."

Why This Matters

The paper concludes that for dangerous environments like Chernobyl, you don't just need a map of where the radiation is; you need to know how much you can trust that map.

The "Dual-Circuit" Idea: Because the "Smart Detective" method is computationally heavy (it takes a lot of brainpower to calculate), the researchers suggest a two-step system:
1. Real-time: Show a quick, rough map immediately so pilots can see where they are.
2. Later: Run the heavy "Smart Detective" calculation to create a highly accurate, high-confidence map for decision-makers.

In short, the paper proves that using advanced math to fill in the gaps of drone data is better than simple guessing, and it's crucial to always show a "confidence meter" alongside the data so people know where the map is reliable and where it's just a guess.

Technical Summary: A Probabilistic Framework for Reconstructing Sparse UAV-Based Radiation Monitoring Data

Problem Statement
Contemporary environmental monitoring, particularly regarding radioactive contamination, faces significant challenges when relying on traditional stationary networks. These systems suffer from low spatial resolution, static coverage, and an inability to rapidly survey hazardous or inaccessible areas. While Unmanned Aerial Vehicles (UAVs) equipped with IoT sensors offer a mobile alternative, they frequently generate spatially heterogeneous and incomplete datasets due to flight trajectory constraints, environmental obstacles, and mission duration.

The core problem addressed is the difficulty of reconstructing continuous environmental contamination fields from these sparse, discrete measurements. Standard interpolation methods often fail to capture complex spatial structures or provide reliable predictions in under-sampled regions. Furthermore, decision-makers require not just a reconstructed map, but an explicit estimation of the confidence (uncertainty) associated with those predictions to assess environmental risk accurately.

Methodology
The study proposes a unified, uncertainty-aware analytical framework that integrates data pre-processing, spatial interpolation, and uncertainty quantification. The methodology is implemented within a hybrid edge-cloud architecture:

System Architecture: A sensor-agnostic system collects data via UAVs, transmitting JSON payloads via MQTT to a cloud server. The architecture supports a dual-circuit approach: real-time visualization of raw data and periodic generation of high-precision analytical models.
Data Pre-processing: To address noise and non-uniformity in airborne gamma-spectrometric data, a two-stage workflow is applied:
- Non-linear Noise Filtering: A median filter is used to remove high-frequency artifacts and outliers (salt-and-pepper noise) while preserving sharp spatial gradients, outperforming linear moving averages.
- Grid-based Aggregation: Data from overlapping flight paths is homogenized by averaging values within virtual grid cells, reducing the dataset to a structurally consistent set of representative points (1,679 points for GPR training).
Spatial Reconstruction Models: Two probabilistic approaches are compared against a baseline:
- Inverse Distance Weighting (IDW): Used as the baseline deterministic method.
- Ordinary Kriging: A geostatistical method utilizing a spherical variogram model to account for spatial correlation.
- Gaussian Process Regression (GPR): A machine learning approach employing a composite kernel (Radial Basis Function + WhiteKernel) to model smooth spatial correlations and measurement noise. A logarithmic transformation is applied to stabilize variance. Hyperparameters are optimized automatically during training.
Validation: The models are validated using an 80/20 random train-test split on a dataset of 3,722 Cesium-137 activity measurements collected in the Chornobyl Exclusion Zone. Performance is evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination ( $R^2$ ).

Key Results
Experimental validation demonstrates that methods accounting for spatial correlation significantly outperform the baseline IDW approach:

Accuracy: Gaussian Process Regression (GPR) achieved the highest performance metrics, reducing RMSE by approximately 23% compared to IDW and 11% compared to Ordinary Kriging. GPR yielded an $R^2$ of 0.96, compared to 0.94 for Kriging and 0.92 for IDW.
Uncertainty Quantification: Unlike deterministic methods, GPR provides spatially explicit estimates of predictive uncertainty (standard deviation). The resulting uncertainty maps clearly visualize regions of high confidence (near flight paths) and low confidence (extrapolated areas), offering a critical tool for risk assessment.
Computational Performance: While GPR offers superior accuracy, it is computationally intensive ( $O(n^3)$ complexity). On the test hardware (Apple M4), fitting the model took approximately 30 seconds for the pre-processed dataset. This necessitates the proposed dual-circuit architecture, where real-time visualization handles immediate needs, while GPR generates high-precision models periodically.

Key Contributions
The paper claims the following primary contributions to the field of environmental monitoring:

Probabilistic Framework: The development of a specific framework for reconstructing sparse UAV-based radiation data using Gaussian Process Regression.
Integrated Uncertainty Quantification: The direct integration of uncertainty estimation into environmental contamination mapping, moving beyond simple deterministic interpolation.
Dual-Circuit Architecture: A demonstration of a system design that combines real-time data visualization with the periodic generation of high-precision probabilistic models, addressing the trade-off between latency and analytical rigor.
Sensor-Agnostic Design: A system architecture capable of integrating diverse IoT sensors (radiological, gas, particulate matter) without altering the underlying processing framework.

Significance and Claims
The authors assert that the proposed framework significantly improves the interpretation of sparse environmental sensing data, particularly under conditions of incomplete spatial coverage. By providing spatially explicit uncertainty estimates, the system enables "confidence-aware environmental decision-making," allowing operators to identify areas requiring additional measurement and to make more conservative risk assessments in high-uncertainty zones.

The study concludes that while Ordinary Kriging remains a reliable and interpretable baseline with lower computational costs, GPR is superior for reconstructing complex spatial fields and quantifying prediction reliability. The work highlights that future environmental monitoring systems must evolve beyond simple spatial interpolation to include probabilistic reconstruction and uncertainty mapping to support effective emergency response and long-term risk assessment. The authors note limitations regarding the single-site study scope and the computational demands of GPR for large-scale, real-time applications, suggesting future work should focus on scalable GPR variants and adaptive UAV path planning.