TLDR: Researchers developed a novel Color-Based Computer Analysis (CBCA) method that integrates image processing and machine learning to accurately determine total oxidant concentration in non-thermal plasma-treated aqueous systems. By analyzing colorimetric shifts in potassium iodide solutions captured by a camera, the method overcomes the subjectivity of traditional titration. Machine learning models, particularly Gradient Boosting and Linear Regression, achieved high predictive accuracy (R² > 0.99), even with a reduced set of color features. This robust, operator-independent approach offers a precise and near real-time alternative for monitoring water treatment efficiency.
Measuring the effectiveness of advanced water treatment systems, particularly those using non-thermal plasma (NTP), has long faced a significant hurdle: accurately determining the concentration of total oxidants. These oxidants, a mix of reactive oxygen and nitrogen species (RONS), are crucial for purifying water, but their fleeting nature and the subjective human element in traditional measurement methods make precise quantification a challenge.
A New Vision for Oxidant Measurement
A groundbreaking study introduces a novel approach called Color-Based Computer Analysis (CBCA) that combines advanced image processing with machine learning to overcome these limitations. This method offers a precise, objective, and real-time way to quantify total oxidant concentration ([Ox]ₜₒₜ) by observing color changes in potassium iodide (KI) solutions.
The Science Behind the Color Change
Non-thermal plasma systems generate various RONS, such as ozone and hydroxyl radicals, which are powerful agents for water, wastewater, and gas treatment. To measure their collective strength, researchers often use the iodometric method, where colorless KI solution reacts with oxidants to form iodine, turning the solution yellow to brown. Traditionally, this color change is then quantified by a process called back-titration, which relies heavily on an operator’s visual perception and can be inconsistent due to factors like fatigue, lighting, and individual eye sensitivity.
How the CBCA Method Works
The CBCA method revolutionizes this process by replacing human observation with a sophisticated automated system. Here’s a breakdown of the steps:
-
Oxidant Production and Video Capture: An NTP system generates oxidants, which are then passed through a KI solution. A custom-built visual data acquisition chamber, named Genesis, equipped with a Raspberry Pi camera, records high-resolution video of the KI solution’s color transitions during oxidation.
-
Chemical Analysis for Reference: Alongside the video recording, traditional titrimetric measurements of [Ox]ₜₒₜ are performed at specific intervals to provide accurate reference data for training the machine learning models.
-
Image Processing Pipeline: The recorded videos are broken down into individual image frames. A robust image processing pipeline is then applied to these frames. This involves several steps: identifying and cropping the region of interest (the KI solution) using deep learning (YOLOv8), converting images into different color spaces (RGB, HSV, Lab), applying masks to isolate key color features, correcting for reflections, enhancing contrast, and removing noise. This meticulous process ensures that only relevant color information is extracted.
-
Data Analysis and Feature Extraction: From the processed images, mean values of color channels (like Red, Green, Blue from RGB; Hue, Saturation, Value from HSV; and Lightness, a-channel, b-channel from Lab) are extracted. These color features are then correlated with the experimentally measured [Ox]ₜₒₜ values. The study found strong linear correlations, particularly with saturation (HSV-S), the blue-yellow axis (Lab-b), and the blue channel (RGB-B), indicating their high responsiveness to oxidant concentration.
-
Machine Learning Model Training: Five different machine learning models—Linear Regression (LiR), Ridge Regression (RR), Random Forest Regressor (RFR), Gradient Boosting Regressor (GBR), and Neural Networks (NN)—were trained and validated using the extracted color features and the corresponding [Ox]ₜₒₜ values. The models were tested with both a full set of nine color features and a reduced set of four highly predictive features (RGB-B, HSV-S, Lab-a, and Lab-b).
Impressive Accuracy and Robustness
The results were highly promising. Models like Gradient Boosting Regressor (GBR) and Linear Regression (LiR) achieved exceptional predictive accuracy, with R² values consistently exceeding 0.99. This means they could explain over 99% of the variance in oxidant concentration based on the color data. Even with a reduced set of four features, the performance remained remarkably high, especially for ensemble models like RFR and GBR, demonstrating their efficiency and robustness.
The CBCA method was rigorously validated against independent experimental data, showing excellent agreement between the model predictions and actual titration measurements. This confirms that the system can reliably predict [Ox]ₜₒₜ with precision comparable to, or even surpassing, traditional manual methods, while eliminating operator bias.
Also Read:
- AI’s Eye on Our Waters: Predicting Pollution with Computer Vision
- Streamlining Methane Detection for Satellites with Machine Learning
The Future of Water Treatment Monitoring
This study marks a significant step forward in analytical chemistry and environmental monitoring. The CBCA method offers a robust, operator-independent, and near real-time solution for quantifying oxidants in aqueous systems. Its adaptability and low-cost optical setup make it suitable for broad adoption in various environmental and industrial contexts where conventional titration is resource-intensive or impractical. Future work aims to extend this image-based machine learning approach to other reaction systems and integrate it into automated control systems for continuous oxidation reactors.
For more details, you can read the full research paper here.
