Rivers transport terrestrial microplastics (MP) to the marine system, demanding cost-effective and frequent monitoring, which is attainable through remote sensing. This study aims to develop and test microplastic concentration (MPC) models directly by satellite images and indirectly through suspended sediment concentration (SSC) as a proxy employing a neural network algorithm. These models relied upon high spatial (26 sites) and temporal (198 samples) SSC and MPC data in the Tisza River, along with optical and active sensor reflectance/backscattering. A feedforward MLP neural network was used to calibrate and validate the direct models employing k-fold cross-validation (five data folds) and the Optuna library for hyperparameter optimization. The spatiotemporal generalization capability of the developed models was assessed under various hydrological scenarios. The findings revealed that hydrology fundamentally influences the SSC and MPC. The indirect estimation method of MPC using SSC as a proxy demonstrated higher accuracy (R2 = 0.17-0.88) than the direct method (R2 = 0-0.2), due to the limitations of satellite sensors to directly estimate the very low MPCs in rivers. However, the estimation accuracy of the indirect method varied with lower accuracy (R2 = 0.17, RMSE = 12.9 item/m3 and MAE = 9.4 item/m3) during low stages and very high (R2 = 0.88, RMSE = 7.8 item/m3 and MAE = 10.8 item/m3) during floods. The worst estimates were achieved based on Sentinel-1. Although the accuracy of the MPC models is moderate, it still has practical applicability, especially during floods and employing proxy models. This study is one of the very initial attempts towards MPC quantification, thus more studies incorporating denser spatiotemporal data, additional water quality parameters, and surface roughness data are warranted to improve the estimation accuracy.
Keywords: Tisza River; hydrological regimes; machine learning; remote sensing; spatiotemporal distribution.