Evaluating traditional empirical models and BPNN models in monitoring the concentrations of chlorophyll-a and total suspended particulate of eutrophic and turbid waters

TLDR: In this article, the authors used in situ sensed reflectance to monitor the concentrations of chlorophyll-a (Chl-a) and total suspended particulate (TSP) of waters in the Pearl River Delta, which is featured by the highly developed network of rivers, channels and ponds, 135 sets of simultaneously collected water samples and reflectance were used to test the performance of the traditional empirical models (band ratio, three bands) and the machine learning models of a back-propagation neural network (BPNN).

Abstract: In order to use in situ sensed reflectance to monitor the concentrations of chlorophyll-a (Chl-a) and total suspended particulate (TSP) of waters in the Pearl River Delta, which is featured by the highly developed network of rivers, channels and ponds, 135 sets of simultaneously collected water samples and reflectance were used to test the performance of the traditional empirical models (band ratio, three bands) and the machine learning models of a back-propagation neural network (BPNN). The results of the laboratory analysis with the water samples show that the Chl-a ranges from 3 to 256 µg·L−1 with an average of 39 µg·L−1 while the TSP ranges from 8 to 162 mg·L−1 and averages 42.5 mg·L−1. Ninety sets of 135 samples are used as training data to develop the retrieval models, and the remaining ones are used to validate the models. The results show that the proposed band ratio models, the three-band combination models, and the corresponding BPNN models are generally successful in estimating the Chl-a and the TSP, and the mean relative error (MRE) can be lower than 30% and 25%, respectively. However, the BPNN models have no better performance than the traditional empirical models, e.g., in the estimation of TSP on the basis of the reflectance at 555 and 750 nm (R555 and R750, respectively), the model of BPNN (R555, R750) has an MRE of 23.91%, larger than that of the R750/R555 model. These results suggest that these traditional empirical models are usable in monitoring the optically active water quality parameters of Chl-a and TSP for eutrophic and turbid waters, while the machine learning models have no significant advantages, especially when the cost of training samples is considered. To improve the performance of machine learning models in future applications on the basis of ground sensor networks, large datasets covering various water situations and optimization of input variables of band configuration should be strengthened.