Integrated workflow for interpretation of satellite imageries using machine learning to assess and monitor algal blooms in Utah Lake, USA

Abstract

An integrated workflow is developed to estimate the spatial distribution of harmful algal blooms, especially cyanobacteria concentrations in inland water bodies. The methodology comprises satellite data extraction and preprocessing for atmospheric and water surface corrections, identifying feature importance, in-situ sample collection, training and testing of machine learning algorithms, and prediction. Six input bands are selected using feature importance algorithms from 12 original bands of Sentinel-2 satellite imagery. In-situ sample data that are synchronous with Sentinel-2 image capture time were obtained from a public database. These models are evaluated and compared using spider plots of different error calculations. The workflow developed in this study and the predicted spatial concentration of harmful algal blooms across the lake can be used to improve warning and advisory systems for the public and avoid exposure. The incorporation of other parameters such as water temperature, nutrient concentrations, and surface wind speed could improve the machine-learning models.