Related Experiment Video
Updated: Jun 21, 2026
09:38
Single-throughput Complementary High-resolution Analytical Techniques for Characterizing Complex Natural Organic Matter Mixtures
Published on: January 7, 2019
Full-spectrum EEM end-member unmixing with statistical validation: A deep learning framework for organic pollution
Jimin Lee1, Soyoung Lee1, Eu Gene Chung1
1Water Environment Research Department, National Institute of Environmental Research, Hwangyong-ro 42, Seogu, Incheon, 22689, Republic of Korea.
Journal of Environmental Management
|June 19, 2026
Summary
This study introduces a new framework using full-spectrum fluorescence EEM images and CNN-unmixing to accurately identify organic pollution sources in mixed watersheds. The method improves upon traditional indices for better source discrimination and pollution load assessment.
Area of Science:
- Environmental Chemistry
- Water Quality Monitoring
- Spectroscopic Analysis
Background:
- Conventional fluorescence-based methods struggle to accurately apportion organic pollution sources in complex, mixed land-use watersheds.
- Traditional approaches condense multidimensional excitation-emission matrix (EEM) spectra into limited indices, hindering source discrimination when multiple pollution sources overlap.
Purpose of the Study:
- To develop and validate an analytical framework for organic pollution source apportionment in mixed land-use watersheds.
- To enhance the tracer capacity of fluorescence spectroscopy for improved source identification and contribution estimation.
Main Methods:
- Utilized full-spectrum EEM images as source-specific end-members to retain detailed spectral information.
- Performed statistical assessments (SSIM, ANOSIM) to evaluate end-member independence and separability.
- Developed a Convolutional Neural Network (CNN)-unmixing architecture to learn source-discriminating patterns and estimate contributions from mixed EEMs.
Main Results:
- Statistical analysis confirmed high within-group similarity and significant between-group separation for eight end-members in EEM spectral space.
- The CNN-unmixing model achieved high accuracy (R² > 0.7, MAE < 0.07) in controlled mixtures, correctly identifying dominant sources and their rankings.
- Field application demonstrated spatial consistency between predicted source contributions and observed pollution patterns, indicating preliminary field-scale applicability.
Conclusions:
- The proposed framework effectively utilizes full-spectrum EEM data and CNN-unmixing for accurate organic pollution source apportionment in mixed land-use watersheds.
- This approach overcomes limitations of traditional fluorescence indices, offering a more robust tool for water quality management and source-priority assessment.