Unveiling the Dynamics of Chlorophyll-a in Arid Oligotrophic Lakes: A Machine Learning Approach in Xinjiang, China

This study focuses on estimating Chlorophyll-a (Chl-a) in nutrient-poor, oligotrophic lakes in Xinjiang, China, from 2020 to 2023. It identifies the limitations of traditional remote sensing methods in detecting low Chl-a levels due to the water's inherent properties. Using MODIS imagery and field data, the research evaluates different models, including empirical, semi-analytical, and machine learning approaches, for Chl-a estimation. Machine learning techniques, particularly the Mixed Density Network (MDN) and Automated Machine Learning (AutoML), proved to be more effective. The research involved comprehensive data collection from 2002 to 2023, including seven detailed field samplings in northern Xinjiang's major lakes. MODIS Aqua data was extensively used, and an atmospheric correction was performed, with the POLYMER method showing high accuracy. The study compared various models for Chl-a estimation, finding that while empirical and semi-analytical models had limitations in accurately quantifying Chl-a, machine learning models, especially MDN and AutoML, showed superior performance. MDN was chosen for its probabilistic nature and balanced error distribution. The model was applied to analyze nine major Xinjiang lakes over two decades. Despite significant pixel loss in daily MODIS images, 8-day Chl-a concentration data products were created using spatiotemporal interpolation. The study revealed spatial and temporal fluctuations in Chl-a concentrations across the lakes, with Jili Lake showing the highest concentration. Overall, the study highlights the importance of advanced machine learning techniques in accurately estimating Chl-a in oligotrophic lakes in arid environments and contributes valuable insights for future environmental monitoring and management in Xinjiang.