Date of Award


Degree Name

Master of Science


Geological Sciences


Hernan A. Moreno


Accurately predicting lake water temperature (LWT) and dissolved oxygen (DO) is crucial for determining threshold values of fish survivability under warmer global conditions, with recreational fishing in reservoirs significantly contributing to regional economies, such as $779 million and $1,891 million annually to the economies of Oklahoma and Texas, respectively. Current mathematical models for temperature and oxygen profiles, which incorporate multi-layer and turbulent mixing equations, are complex and challenging to parameterize, particularly due to uncertainties in acquiring sufficient data for training and validation. Leveraging the flexibility and information extraction power of machine learning (ML) methods, this master thesis aimed to set up and test ML and deep learning (DL) models to predict LWT and DO across 12 lakes within the Red River Basin of the South in the United States, using historical spatially distributed measurements. Five ML approaches, including Random Forest (RF), Gradient Boosting Extreme (XGBoost), Tree-Boosting with Gaussian Process and Mixed Effects Model (GPBoost), Support Vector Machine (SVM), and Deep Learning (DL), were assessed using numerical k-fold cross-validation metrics. The results highlight GPBoost as the most effective method for predicting LWT and DO, which is attributed to their incorporation of interpretable physical variables. Notably, GPBoost exhibited robust performance under various lake conditions, while RF, XGBoost, and SVR showed signs of overfitting. Comparisons with traditional 1-D numerical approaches underscore the potential of ML algorithms for faster and more precise results, offering valuable insights into the dynamics of lake ecosystems and emphasizing the need for alternative methods to capture their complexities effectively.




Received from ProQuest

File Size

100 p.

File Format


Rights Holder

Isabela Suaza Sierra