|
An autonomous water quality monitoring system using computer vision and machine learning |
|---|---|
| รหัสดีโอไอ | |
| Title | An autonomous water quality monitoring system using computer vision and machine learning |
| Creator | Nguyen Khac Vinh Khang |
| Contributor | Warut Pannakkong, Advisor |
| Publisher | Thammasat University |
| Publication Year | 2568 |
| Keyword | Water quality monitoring, Computer vision, Machine learning, Digital image analysis, Wastewater classification, Chemical oxygen demand |
| Abstract | Access to reliable chemical oxygen demand (COD) monitoring remains a challenge in resource-constrained settings, where conventional spectrophotometric analysis requires costly laboratory equipment and specialized operation. This study presents a multi-stage vision-based framework that combines smartphone imaging, automated region-of-interest localization, environment classification, COD class classification, engineered color feature extraction, and machine learning regression for COD estimation. A controlled dataset was created from COD samples spanning ultra-low to high-range concentrations, and images were acquired under varied lighting and zoom conditions using a consumer-grade smartphone camera. A MobileNetV3-based classifier was used to predict the acquisition environment, while ROI detection and feature engineering transformed vial images into structured tabular descriptors derived from RGB, HSV, Lab, and perceptual ΔE-based features. For COD class classification, gradient-boosting classifiers were evaluated to assign samples into UL, LR, HR, and HR+ ranges. Multiple regression models, including linear, tree-based, boosting, and ensemble methods, were then optimized and compared using class-wise evaluation, error metrics, and inferential statistics. Results show that ensemble regressors achieved the strongest performance in the UL, LR, and HR classes, while prediction error increased substantially in the HR+ range. Statistical analysis further confirmed that COD class has a much stronger influence on error than model choice alone. Overall, the proposed framework demonstrates that low-cost smartphone imaging, combined with class-aware machine learning and statistical validation, can provide a reproducible and scalable alternative for COD monitoring in resource-limited environments. |
