Remote estimates of suspended particulate matter in global lakes using machine learning models

• DOI: 10.1016/j.iswcr.2023.07.002
• Publication Date: 2023-07-16T12:03:04Z
• Authors: Zhidan Wen; Qiang Wang; Yue Ma; Pierre Andre Jaci...; Ge Liu; Sijia Li; Yingxin Shang; Hui Tao; Chong Fang; Lili Lyu; Baohua Zhang; Kaishan Song
• Keywords: Gradient boosting

Suspended particulate matter (SPM) in lakes exerts strong impact on light propagation, aquatic ecosystem productivity, which co-varies with nutrients, heavy metal and micro-pollutant waters. In lakes, SPM absorption backscattering, ultimately affects water leaving signals that can be detected by satellite sensors. Simple regression models based specific band or hand ratios have been widely used for estimate the past moderate accuracy. There are still rooms model accuracy improvements, machine learning may solve non-linear relationships between spectral variable We assembled more than 16,400 situ measured from six continents (excluding Antarctica continent), of 9640 samples were matched Landsat overpasses within ±7 days. Seven algorithms two simple methods (linear partial least squares models) to performance compared. To overcome problem imbalance datasets regression, a Synthetic Minority Over-Sampling technique Gaussian Noise (SMOGN) was adopted this study. Through comparison, we found gradient boosting decision tree (GBDT), random forest (RF), extreme (XGBoost) demonstrated good spatiotemporal transferability SMOGN processed dataset, has potential map at different year quality land surface reflectance images. all tested modeling approaches, GBDT accurate calibration (n = 6428, R2 0.95, MAPE 29.8%) collected 2235 across world, validation 3214, 0.84, 38.8%) also exhibited stable performance. Further, performances RF (R2 0.93) 0.86, 24.2%) datasets. applied typical satisfactory result obtained. addition, evaluated historical measurements coincident sensors (L5-TM, L7-ETM+, L8-OLI), thus monitoring temporal variations, tracks lake dynamics approximately four decades (1984–2021) since Landsat-5/TM launched 1984.