XGBoost-Based Predictive Modeling Using Sentinel-2 Satellite Imagery and Empirical Data of Jakarta River Water Quality Rijaldi R.M. (1), Prayoga G. (1)*, Faskayana Y.S. (1,2), and Firmansyah F.S. (1)
1) Center for Environmental Research, IPB University, Indonesia
2) The Graduate School of Agriculture, Tokyo University of Agriculture and Technology, Japan
*gatotprayoga16[at]gmail.com
Abstract
Machine learning methods have shown strong potential in estimating water quality parameters. Satellite remote sensing provides an efficient approach for monitoring water quality over extended periods and spatial variations, thereby reducing reliance on resource-intensive field monitoring. Therefore, this study aims to develop a predictive approach using XGBoost regression and Sentinel-2 satellite imagery to estimate river water quality. Dataset from 60 river monitoring stations across 12 river sections, collected over 16 periods between 2021 and 2024, was used as empirical data and then paired with harmonized data from Sentinel-2. The dataset paired in-situ measurements with reflectance values from relevant Sentinel-2 bands for model training. Using Google Earth Engine (GEE), reflectance values from Sentinel-2 bands were extracted for each sampling point and acquisition date. The derived spectral features were then used to train XGBoost regression models for each water quality parameter. The models showed moderate predictive performance, with R square values of 0.65 (turbidity), 0.61 (TSS), 0.58 (TDS), 0.63 (color), and 0.67 (transparency). Corresponding RMSE values were 8.7 NTU (turbidity), 45.2 mg/L (TSS), 60.4 mg/L (TDS), 25.8 Pt-Co units (color), and 6.3% (transparency), indicating a fair level of accuracy with room for improvement across parameters. This study demonstrates the possibility of estimating river water quality using remote sensing. This approach could enhance the effectiveness and efficiency of water quality management, particularly in challenging and remote areas. However, several significant limitations, such as the limited availability of Sentinel-2 imagery that coincided precisely with the in situ sampling dates, may have introduced some temporal discrepancies in the dataset.
Keywords: water quality parameters- remote sensing- machine learning- XGBoost- Jakarta
