Inland water's trophic status classification based on machine learning and remote sensing data

• Autor(a) principal: Watanabe, Fernanda S.Y. [UNESP]
• Data de Publicação: 2020
• Outros Autores: Miyoshi, Gabriela T. [UNESP], Rodrigues, Thanan W.P., Bernardo, Nariane M.R. [UNESP], Rotta, Luiz H.S. [UNESP], Alcântara, Enner [UNESP], Imai, Nilton N. [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.rsase.2020.100326; http://hdl.handle.net/11449/200465
• topic: Artificial neural network; Multispectral data; Random forest; Remote sensing; Support vector machine

In this work, we tested machine learning algorithms in classifying waters in a reservoir cascade with basis in trophic state. The classification was done through remote sensing reflectance (Rrs) measurements collected in situ. Chlorophyll-a (chla) content determined in the laboratory were used to define the trophic state in the sampling points distributed in four reservoirs (Barra Bonita, Bariri, Ibitinga and Nova Avanhandava), located at the Tietê River, Brazil. Those four impoundments exhibit widely differing optical properties from each other, which is rather evident in relation to chla concentration. From the dataset collected in the reservoir cascade, a trophic gradient is observed, decreasing from up-to downstream. To classify the trophic state, we tested three machine learning algorithms: Artificial Neural Network (ANN), Random Forest (RF) and Support Vector Machine (SVM). Results showed that ANN and RF algorithms exhibited the best performance in classifying the different trophic state in the cascade of reservoirs. Both approaches raised a global accuracy of 80.00% and average area under Receiver Operating Characteristics (ROC) curve (AUCROC) of 0.928 and 0.912, respectively. Comparing the machine learning approaches with a parametric algorithm, only SVM presented a slightly lower performance. The outcomes of this classification can be useful for trophic state mapping considering the large cascade of reservoirs or rivers. In addition, it can give a direction in bio-optical modeling studies, which have shown that a unique bio-optical algorithm is unable to accurately retrieving concentrations of optically active constituents in aquatic system with high optical variability. So that, it is possible to develop specific chla prediction models considering the optical characteristics of each stretch of river, since machine learning-based classifications (ANN and RF) indicate different optical regions.