Spatial and Temporal Analysis of Water Quality in Southeastern Louisiana Using Machine Learning and Google Earth Engine

Water quality is a critical component of the biogeophysical environment. The spatial and temporal variation of this part of the environment is crucial for stakeholders. Remote sensing offers a unique opportunity to estimate the water quality for non-point scenarios. This dissertation evaluates how pixel window matching affects the analysis of water quality parameters retrieval. The study focuses on how window pixel matching impacts different machine learning algorithms and examines the seasonal effect of pixel matching on the retrieval of water quality parameters for different sensors. The first chapter implemented a support vector regressor for total suspended solids (TSS) and chlorophyll-a during different seasons (dry and wet months). The root mean square error (RMSE) for TSS is slightly better for wet months compared to dry months by 15% on average under similar window matching scenarios. In comparison, chlorophyll-a decreases by 8% from dry months to wet months. The second chapter implemented 2-D convolutional neural network regression for TSS to explore the tradeoff between pixel window matching and different patch sizes (9, 15, and 32). Patch 32 was better under the 5-day window. The third chapter explored the effect of the enhanced temporal resolution of Geostationary Operational Environmental Satellites (GOES) under window matching settings of 1, 2, 3 and 4 days. In this chapter, GOES images, which have the closest match-up to the ground truth for TSS and chlorophyll-a, showed the lowest RMSE of 3.67 ppm and 4.2 mg/L, respectively, and the RMSE increased with increased window match-up. The fourth chapter applied seasonal trend decomposition using Loess (STL) to examine seasonality in the GOES estimated three-year series of TSS and chlorophyll-a for four sites on Lake Pontchartrain, Louisiana. The STL model found an increasing trend for chlorophyll-a and a decreasing trend for TSS for all sites. The study concluded that pixel match-up could provide a unique opportunity in environmental study and modeling irrespective of machine learning algorithms, sensors and seasons. This study has demonstrated that machine learning algorithms can benefit from sensors for different water quality parameters and showed how different seasonal scenarios can impact the retrieval of water quality parameters.