Deep learning-based classification of nitrate and nitrite concentrations from water samples using colorimetric test...

Area of Science:

• Environmental Science
• Computer Science
• Analytical Chemistry

Background:

• Accurate water nitrate and nitrite monitoring is crucial for agriculture, public health, and aquatic ecosystems.
• Traditional testing methods are precise but impractical for frequent, on-site analysis due to cost and complexity.

Purpose of the Study:

• To develop and evaluate deep learning-based computer vision techniques for classifying nitrate and nitrite concentrations in water using images of colorimetric test strips.
• To compare the performance of deep learning models against classical machine learning approaches.

Main Methods:

• Images of colorimetric test strips were acquired using an RGB IMX219 camera under controlled illumination.
• Deep learning models (MLP, AlexNet, VGG16, ResNet18, GoogLeNet) and classical machine learning baselines were trained and evaluated.
• Stratified fivefold cross-validation and an independent test set were used for performance assessment.

Main Results:

• For nitrate, ResNet18 and GoogLeNet achieved near-perfect 100% test accuracy, significantly outperforming classical methods (max 83.5% test accuracy).
• For nitrite, GoogLeNet demonstrated the highest performance with 97.48% test accuracy, surpassing the best classical model (max 83.19% test accuracy).
• Deep Convolutional Neural Network (CNN) based feature learning showed a significant performance advantage.

Conclusions:

• Deep learning computer vision is highly effective for rapid, image-based water quality assessment of nitrate and nitrite.
• The proposed system offers a promising, accurate, and potentially cost-effective alternative to traditional water testing methods.
• Further evaluation under real-world conditions is warranted to assess broader deployment suitability.