Assessment of Inland Lake Water Color and Dissolved Organic Content via Satellite-Based Machine Learning
Board Location: #31
Discipline: Ecology Environmental and Earth Sciences
Subcategory: Water
Session: 3
Aisha Malik - CUNY Hunter College
Co-Author(s): Fahmeda Khanom, CUNY New York City College of Technology, New York City; Touheda Khanom, CUNY New York City College of Technology, New York City; Marzi Azarderakhsh, CUNY New York City College of Technology, New York City; Hamidreza Norouzi, CUNY New York City College of Technology, New York City; Reginald Blake, CUNY New York City College of Technology, New York City
In the 21st century, climate change stands as a formidable threat to New York State’s natural resources, including over 3000 lakes and their watersheds. The lakes within the Adirondack Park have benefited from strict land use laws due to the Clean Air Act to control the anthropogenic impacts on the lakes’ health. In the late 1970s, the Adirondack Lake Survey (ALS) took extensive field data including water chemistry to better understand the health of more than 30 represented lakes to monitor the recovery from acidification. The impact of climate change and anthropogenic activity raises concerns about algal blooms, rising temperatures, aquatic habitats, and other threats to the lake ecosystem. However, there is a significant data gap in monitoring beyond the selected lakes.In this study, the spatial and temporal patterns of water chemistry data are explored by leveraging current and long-term monitoring programs, and by integrating satellite remote sensing imagery to inform future research on watercolor change as an indicator of water quality beyond the selected sampling lakes through a future Survey of Climate change and Adirondack Lake Ecosystems (SCALE) Pilot Program. A dataset with sampling since 1985 is compiled for 50 lakes and Dissolved Organic Content (DOC) is selected for comparison with several empirical algorithms using Landsat 5 & 8 surface reflectance (SR) observations.While many studies develop predictive relationships between remotely sensed surface reflectance and water parameters, these relationships are often limited to a specific geographic region and have little applicability in other areas. To monitor DOC remotely, region-based relationships must be developed. The preliminary data analysis of several empirical algorithms does not show a strong correlation for the represented ALTM lakes. However, they exhibit consistent long-term trends using Landsat-5 SR data suggesting lake color change in several of the sampled lakes. Since different empirical algorithms performed differently for various lakes, a machine-learning approach that can learn the complex relation between the inputs and data types is applied. Several machine-learning techniques, including boosting and bagging, are employed to estimate DOC using different input features from satellite surface reflectance data. The model utilizes 70% of the data from each lake selected for training and 30% for testing performance. To find the best-performing model, we examined the impact of lake classification, atmospheric correction algorithms, and lake water depth on the model’s performance. This analysis is performed on an openly accessible Python script on the Google Earth Engine Platform for processing cloud-based publicly available satellite observation data and will allow the determination of relationships over various lakes and studying the impact of climate change within the larger Adirondack region.
Funder Acknowledgement(s): This research project was supported by the NSF GRANT# AGS-1950629 (NSF-REU).
Faculty Advisor: Marzi Azarderakhsh, MaAzarderakhsh@citytech.cuny.edu
Role: I collaborated with my team by creating, investigating, and analyzing empirical and machine learning algorithms to determine the most effective models that estimate and predict accurate CDOM and DOC concentrations in inland lakes using satellite imagery. I aided in coding an automated Python algorithm that spans over 50 lakes using Adirondack Long-Term Monitoring (ALTM) data. This algorithm produced Landsat-5 & 8 maps and graphs that depict CDOM and DOC time-series/scatterplots, reflectance data graphs using satellite bands for desired intervals in time to analyze seasonal variation, and tested machine learning models. I’m currently the first author of an in-progress paper on this research.