Evaluation of Parametric and Nonparametric Algorithms for the Estimation of Suspended Particulate Matter in Turbid Water using Gaofen-1 Wide Field-of-view Sensors

• Published in: Journal of the Indian Society of Remote Sensing Vol. 49; no. 11; pp. 2673 - 2687
• Main Authors: Liu, Yao, Fan, Jie-Ping, Jiang, Hui
• Format: Journal Article
• Language: English
• Published: New Delhi Springer India 01.11.2021; Springer Nature B.V
• Subjects: Algorithms; Artificial neural networks; Back propagation; Back propagation networks; Earth and Environmental Science; Earth Sciences; Evaluation; Field of view; Lakes; Machine learning; Monitoring; Neural networks; Nonparametric statistics; Particulate emissions; Particulate matter; particulates; Radial basis function; reflectance; regression analysis; Regression models; Remote sensing; Remote Sensing/Photogrammetry; Research Article; Retrieval; Satellite imagery; Sensors; Spatial resolution; Spectra; spectral analysis; Spectral reflectance; Spectral sensitivity; Turbidity; Poyang Lake; Gaofen-1; Neural network; Turbid water; Suspended particulate matter; Poyang Lake
• ISSN: 0255-660X; 0974-3006
• DOI: 10.1007/s12524-021-01405-7

Quantifying the concentration of suspended particulate matter ( C SPM ) is necessary for the evaluation of the ecological processes, matter transfer, and environments in lakes. Traditional monitoring via the use of field measurements lacks the spatial coverage necessary for detailed analysis in large-area lakes. In contrast, the macroscopic, real-time monitoring of C SPM can be achieved using data from remote-sensing satellites. Due to the major limitations of the existing analytical and semi-analytical algorithms, namely the lake locations and limited data, regression algorithms have become important tools for the analysis of the C SPM values in lakes. The purpose of this study is the evaluation of both parametric and nonparametric regression algorithms for C SPM estimation based on the simulation of Gaofen-1 (GF-1) wide field-of-view (WFV) satellites for the imaging of turbid water. A total of 71 samples collected during four cruises were analyzed to determine the spectral reflectance and C SPM in Poyang Lake. The results indicate that the spectral ratio (SR) parametric regression model and the extreme learning machine (ELM) nonparametric regression model based on in situ SPM and spectral data perform relatively better than the other investigated models. A SR fitting model using band3 ( λ  = 660 nm)/band2 ( λ  = 555 nm) of the GF-1 WFV sensor was established to estimate the C SPM value. The evaluation results indicate that the SR model achieved a high retrieval accuracy ( R 2  = 0.901, RMSE = 8.22 mg·L −1 ); this is because the model was more sensitive and could effectively weaken, and even partially eliminate, the effects of external factors. The ELM model achieved a higher retrieval accuracy ( R 2  = 0.903, RMSE = 8.05 mg·L −1 ) than the other neural network-based algorithms, namely the backpropagation (BP), radial basis function (RBF), and cascade-forward neural network (CFNN) algorithms. The ELM model is well suited for both low and high turbidity, and has a robust ability to accurately retrieve the C SPM value. In Poyang Lake, the C SPM values were found to have a spatial characteristic and were higher in the northern waters than in the southern waters. Overall, this study demonstrates the applicability of GF-1 WFV sensors with an advanced spatial resolution and an increased spectral sensitivity for monitoring the C SPM values in high-turbidity water, and the ELM model is recommended for application to GF-1 images to satisfactorily produce quantitative C SPM maps.