| Summary: | The incidence of harmful algal blooms (HABs) caused by blue-green algae has been increasing in coastal and freshwater ecosystems of the United States in recent years, and has had great influence on ecosystem, economic, and public health. This thesis aims at testing the feasibility of using machine learning methods in comparison to traditional regression models to detect and map the blue-green algae distribution in low-medium biomass waters (Chl-a < approx. 20 μg/L) from a Landsat 8 image with the support of some in situ Chl-a measurements in Harsha Lake, Ohio. Two algorithms were compared: one is the conventional empirical method Stepwise Multiple Linear Regression to see if there is a strong linear relationship between measured Chl-a concentrations and the Landsat 8 spectral data in the study area, and the other is one of the most popular machine learning methodsRandom Forests. Major findings include: (1) both a conventional linear regression model and a Random Forests model worked well in mapping the extent and biomass of blue-green algae in Harsha Lake on September 21, 2015, but the Random Forests model outperformed the linear regression model; (2) the prediction surface from the Random Forests method illustrated that 89.30% of Harsha Lakes area had Chl-a values less than 10 µg/L on the sampling date, while only 10.70% of the entire study area had Chl-a concentrations between 10 µg/L and 20 µg/L. Higher Chl-a values (especially for Chl-a larger than 10 µg/L) were mostly distributed in the mouths of rivers or streams, which might be caused by the influx of nutrients from agricultural or urban land use by rivers and streams. The results show the utility of the Random Forests approach based on Landsat 8 imagery in detecting and quantitatively mapping low biomass HABs, which is considered to be a challenging task.
|
|---|
