A big data-driven model for water quality estimation with high spectral and spatial resolution imagery

Abstract

Hyperspectral imagery is effective to identify harmful cyanobacteria blooms by having an advantage in accurate cyanobacteria detection with high spectral and spatial resolution. In addition, the hyperspectral image contains enormous amount of optical information in every single pixel. Thus, based on the big data imageries, this research utilized deep learning models called stacked autoencoder and convolutional neural network to estimate phycocyanin and chlorophyll-a concentrations, and produced the pigment maps. Stacked autoencoder (SAE) and point-centered regression CNN (PRCNN) models were developed with hyperspectral imagery input and they showed more precise PC and Chl-a estimated result than the results from conventional algorithms. In addition, the PC and Chl-a concentration maps from SAE and PRCNN showed reasonable pigment concentration levels by significantly tracing the actual spatial distribution of the pigments. Thus, this research demonstrated that deep learning approaches had the strong capacity for detection and quantification of harmful cyanobacteria with high accuracy.