Harnessing Machine Learning and Citizen Science Data to Improve Remotely Sensed Estimates of Lake Water Clarity and Document Regional Clarity Trends

Freshwater lakes and reservoirs constitute a small percentage of the Earth's surface but play a disproportionately large role in providing important ecosystem services to humans. Lakes with clear water are particularly valuable amenities that contribute to local and regional economies. Water clarity (as Secchi depth) has been measured routinely by citizen scientist volunteers in New York, Vermont, New Hampshire, and Maine, creating a unique data set of >150,000 measurements that span over 30 years and include more than 500 lakes. Using this data set, we are working to improve optical remote sensing estimates of water clarity in lakes across the four-state region. After extracting satellite data for lakes >150 hectares in area from Landsat 4, 5, 7 and 8 with Google Earth Engine, we used a Random Forest approach to fit models relating the in-situ clarity observations to the optical satellite bands and several other variables. Satellite-based estimates of water clarity across a range of Secchi depths using this approach generated lower mean absolute and root mean square error when compared to previously published algorithms. Overall, our approach estimated Secchi depth to within 1.36 meters over all Secchi depths and to within 0.3 meters for Secchi depths <3 meters; this study represents the largest lake water clarity validation study to date. In addition, we used our new algorithm to hindcast changes in lake clarity across over 10,000 lakes in our four-state region. Our results suggest that the combination of citizen science data, Google Earth Engine, and machine learning methods provides a flexible framework to extend monitoring of lake water quality characteristics across large regions and generate data sets that are important to aquatic ecologists, water resource managers, and interested communities.