A convolutional neural network regression for quantifying cyanobacteria using hyperspectral imagery

Abstract

Remote sensing is useful for detecting and quantifying cyanobacteria blooms for managing water systems. In particular, airborne hyperspectral remote sensing has an advantage in precise cyanobacteria detection with high spatial and spectral resolution. Many bio-optical algorithms have been developed and utilized to estimate algal concentration. However, achieving the optimal conventional optical model accuracy is still challenging in freshwater owing to the biophysical complexity of the inland water and the seasonal reflection of site-specific optical properties. Thus, this study applied convolutional neural network (CNN) with various input windows to estimate the concentrations of phycocyanin (PC) and chlorophyll-a (Chl-a), and generated a phytoplankton pigment map. We proposed that the Point-centered regression CNN (PRCNN) showed accurate PC and Chl-a simulations, with R2 > 0.86 and 0.73, respectively, and root mean square errors of <10 mg·m−3, which were smaller than the conventional optical algorithm in our study area. In addition, the generated PC and Chl-a map from PRCNN closely followed the spatial distribution of the pigment and showed reasonable concentration levels. Through testing we found that a small input size and deep spectral bands contributed to the CNN model to achieve strong capacity to reflect the dynamic spatial feature of phytoplankton pigments. Therefore, this study demonstrated that CNN regression has the potential to detect and quantify cyanobacteria with high accuracy and can be an alternative to bio-optical algorithms.