Near real-time water quality monitoring with high resolution satellite images and machine learning methods | |
| Author | Shirodkar, Abhishek Babuso |
| Call Number | AIT Thesis no.WM-25-02 |
| Subject(s) | Water quality--Measurement Water--Remote sensing Machine learning |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Water Engineering and Management |
| Publisher | Asian Institute of Technology |
| Abstract | Timely water quality (WQ) monitoring is crucial to the management of inland aquatic systems, especially in urbanizing basins like Thailand's Chao Phraya. Field-based observations, while effective, are sparsely covered and reported delays. This study propose a near real-time monitoring method integrating machine learning (ML) and high-resolution Landsat 8/9 satellite imagery to predict eight WQ parameters: turbidity, total dissolved solids (TDS), electrical conductivity (EC), salinity, temperature, dissolved oxygen (DO), total phosphate (TP), and pH. Spectral features were extracted by means of Pearson correlation from coordinated in-situ data with satellite overpasses. Strong correlations were found between EC, TDS, and salinity and B2/B4 (blue-to-red) band ratio; turbidity significantly correlated with NDTI (r = 0.52) and B3/B4 (green to-red) negatively. DO was marginally correlated to SWIR bands, while TP and pH correlated to linear band combinations of visible and NIR wavelengths. Three ML models; Support Vector Machine (SVM), Random Forest (RF), and Classification and Regression Trees (CART), were developed and validated. RF outperformed the others on optically active parameters with R² values of 0.81 (turbidity), 0.75 (TDS), and 0.78 (EC), and moderate performance for salinity (0.60) and temperature (0.34). For non optical parameters, a hybrid combination of optical predictors enhanced RF performance, with resultant R² scores of 0.75 (DO), 0.72 (TP), and 0.51 (pH). CART indicated overfitting, whereas SVM outperformed overall. Spatial and seasonal predictions at ungauged locations based on 60 Landsat scenes from the 2024 period of monitoring were gathered. Monthly prediction composites suitably represented WQ dynamics well and were analyzed against Thailand standards of Class 2 and Class 3. Turbidity reported substantial wet-season degradation, often crossing Class 3. DO and TDS were generally in Class 2; EC and salinity in fluctuation between Class 2–3. TP and temperature were on the margins, suggesting nutrient enrichment and climatic stress. Despite minor underestimation for certain parameters, the RF model demonstrated consistent robustness and scalability. This study offers a reproducible and cost-effective regional WQ monitoring framework and enables real-time management programs across Southeast Asia's data-scarce zones using open-access Earth observation and ML resources. |
| Year | 2025 |
| Type | Thesis |
| School | School of Engineering and Technology |
| Department | Department of Civil and Infrastucture Engineering (DCIE) |
| Academic Program/FoS | Water Engineering and Management (WM) |
| Chairperson(s) | Shanmugam, Mohana Sundaram |
| Examination Committee(s) | Shrestha, Sangam;Natthachet Tangdamrongsub |
| Scholarship Donor(s) | Global Water and Sanitation Center (GWSC);AIT Scholarship |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2025 |