Advancing lakes algal chlorophyll estimation in the contiguous USA: A comparative study of machine learning models and satellite data

Algal blooms are ubiquitous in lentic ecosystems and pose a risk to human and other organisms' health. Accurate measurement of chlorophyll-a (CHL-a) in lakes at a macroscale is challenging due to the optical complexity of individual water bodies, which hinders the optimization of conventional bio-optical algorithms. This study harnesses the synergy of satellite remote sensing and machine learning (ML) to enhance CHL-a quantification from space. Given the cost and logistical demands of in-situ CHL-a data collection, especially over vast areas, we explore the potential of the open-source AquaSat dataset for CHL-a estimation across the contiguous USA. We assess the performance of four ML algorithms (random forest, extra tree regressor, bagging regressor, and xgboost model), discern the most influential spectral bands and indices, and compare these methods to established remote sensing techniques for CHL-a prediction. Both bagging regressor and random forest performed equally well on all AquaSat data or data from each sensor separately (R2 = 0.35-0.54, RMSE = 20.48-23.90 mu g/ L). Model-agnostic SHAP summary plots were used to identify important indexes in CHL-a estimation. Spatiotemporal validations demonstrated the models' reliability across diverse conditions, with better generalizability in spatial domains compared to seasonal or yearly transitions. The accuracy of algorithms for estimating CHL-a depends on the satellite sensor. We found that by comparing remote sensing studies with various atmospheric correction approaches, the Landsat collection 1 (LC1) surface reflectance product offers consistent CHL-a estimates throughout the USA. Overall, acknowledging the existing limitations and challenges of such approaches, this research illustrates the potential of utilizing open-source data with ML to facilitate large-scale estimation of lake CHL-a.