Xin Lan

Xin Lan is a PhD student in the Department of Geography, Environment, and Spatial Science, with a dual major in the Environmental Science and Policy Program. He earned his Master's degree in Earth and Environmental Engineering from Columbia University. There, he cultivated a multifaceted research interest in water resources management by melding environmental science, specifically hydrology, with policy. Currently, he's delving into research on the historical trends of water coverage and temperature fluctuations in the U.S., utilizing tools like Google Earth Engine, remote sensing techniques, and machine learning models.
DISSERTATION TITLE	UNDERSTANDING THE THERMAL CHANGES IN FRESHWATER LAKES THROUGH OBSERVATIONS, PHYSICAL MODELS, AND MACHINE LEARNING TECHNIQUES
DATE	Friday, April 3, 2026
TIME	3:00 PM - 6:00 PM EDT
LOCATION	Geography 105
COMMITTEE MEMBERS	Dr. Lifeng Luo (Chair, Geography and ESPPP) Dr. Julie Winkler (Geography) Dr. Pang-Ning Tan (Computer Science and Engineering) Dr. Yadu Pokhrel (Civil and Environmental Engineering)

Lakes are critical freshwater ecosystems where water temperature regulates fundamental ecological processes, including metabolic rates, nutrient cycling, and oxygen availability throughout the water column. As climate change increasingly alters lake thermal structure, understanding and predicting vertical temperature dynamics becomes essential for managing freshwater ecosystems. This dissertation investigates vertical lake water temperature dynamics through three progressive studies spanning observational analysis, process-guided ensemble deep learning development, and regional-scale application across the Midwestern United States.

The first study analyzes 17 years of in situ multi-depth observations from Seneca Lake, New York, to characterize vertical warming trends and quantify the contributions of climate and anthropogenic drivers, and further leverages remotely sensed surface water temperature data to assess the spatial impact of thermal pollution from a nearby power plant. Results from linear regression and the Theil-Sen estimator reveal significant warming extending below the lake surface, with principal component and contribution analyses indicating that climate changes, particularly air warming, are more critical than anthropogenic factors in explaining long-term temperature patterns, although land cover change can amplify climate-driven warming. The study further found that the adverse effects of thermal pollution are primarily confined to the area adjacent to the power plant. These findings highlight the importance of multi-depth temperature monitoring for understanding lake thermal responses, yet such long-term observations remain scarce across most lakes, motivating the development of modeling approaches for predicting lake temperature profiles.

The second study develops a process-guided ensemble deep learning framework using Lake Mendota, Wisconsin, integrating four sequential deep learning architectures with simulation-based pretraining, depth-wise ensemble learning, and energy conservation constraints, and systematically evaluates the effectiveness of each component across varying levels of data availability. Among the individual architectures, Attention-LSTM and CNN-LSTM showed greater improvement in depth wise prediction under sparse observations, while simulation-based pretraining and the depth-wise ensemble produced more stable and reliable temperature profiles across all architectures. The energy conservation constraint provided only marginal additional benefit, suggesting that integrating physical knowledge through model initialization via pretraining is more effective than imposing it as a soft constraint during ensemble training. While this framework demonstrates effectiveness for a single well-monitored lake, extending it to regional scales requires incorporating diverse data sources to characterize lakes across broad geographic areas.

The third study extends this framework to 806 lakes across 10 Midwestern states to predict daily vertical temperature profiles during ice-free seasons from 2017 to 2021. The framework processes 30-day North American Land Data Assimilation System meteorological sequences through four parallel temporal encoders, with Feature-wise Linear Modulation (FiLM) adapting the learned representations to lake-specific conditions using HydroLAKES morphometric attributes and AlphaEarth satellite-derived watershed embeddings. The encoders are pretrained on General Lake Model simulations to initialize physically consistent thermal dynamics before fine-tuning with in situ observations. The FiLM-adapted predictions from the four encoders are combined through learned depth-wise ensemble weights that dynamically allocate each architecture’s contribution at different depths. The model was fine-tuned and evaluated using 2,675 lake-date observations, with 82 spatially independent lakes reserved for testing, achieving a test RMSE of 2.18°C without lake specific calibration. SHapley Additive exPlanations analysis revealed that meteorological variables dominate surface temperature predictions, while static lake properties and satellite-derived land surface characteristics become increasingly important at greater depths. The resulting gap filled dataset provides spatio-temporally complete temperature records for a region where most lakes lack continuous monitoring, demonstrating that this framework can produce accurate and interpretable predictions for understanding lake thermal dynamics and ecosystem management