Student Projects

A Machine Learning Approach to Synoptic-Scale Waterspout Prediction in the Florida Keys by Jonny Benoit

Machine learning and statistical methods are used to predict waterspout occurrence in the Florida Keys from a global weather model (NCAR). The results here match the skill of previous regression models that used in situ data from Key West. Bayes Theorem is used to determine the significance of the new waterspout prediction method and convert model outputs to waterspout probability. Conducting machine learning on a pixelwise basis indicates that the pixels near the Florida Keys are most predictive of waterspout occurrence, and ensemble methods are used to increase predictive power. [report]

Predicting Streamflow Using Machine Learning in Northern Canada by Sarah Esenther & Ekaterina Lezine

Predicting streamflow under varying climate conditions is critically important for managing water resources. This is especially true as climate change intensifies the hydrologic cycle in certain regions, such as the northern latitudes. In the past, physical models have been used to predict streamflow based on complex equations describing how water moves across a landscape. These models are difficult to implement and require large amounts of data for validation that is not always available, especially in remote, northern watersheds. Machine learning offers a new way to predict streamflow without significant expertise with physical modeling. In this study, we demonstrate the utility of three different machine learning methods - random forest, multilayer perceptron, and long short term memory – to predict streamflow for three different river gauges in northern Canada. We find that none of the models produce consistently accurate discharge predictions across all stations. We emphasize the need for larger training datasets, better training data, and more precise tuning methods to build more accurate and robust models. [report] [code]

Identifying Moulins on the Greenland Ice Sheet using Ice Sheet Surface and Bedrock Topology by Carolyn Lober & Emma Perkins

Vertical drainage columns in the ice sheet called moulins make up a significant portion of surface water drainage on the Greenland Ice Sheet. Here, digital elevation model (DEM) data were used to train a model to identify moulins based on topology of the ice sheet surface (Greenland Ice Mapping Project/GIMP) and bedrock (IceBridge BedMachine Greenland/BedMachine). Features used were elevation, aspect, slope, and planform and profile curvature for both GIMP and BedMachine; moulin data labels used were from a Landsat-based mapping study. All were sampled at a 30m pixel size. Undersampling of non-moulin pixels was used in combination with oversampling of moulins with SMOTE to attempt to account for an extremely unbalanced dataset (only 872 moulins in approximately 41 million pixels). All combinations of resampling and algorithms resulted in either an extremely high rate of false positives or no predicted moulins at all. Among all of these poor-performing models, the weighted logistic regression on data with random undersampling of the majority class performed the best on the validation data with an overall model accuracy of 0.0898, a true positive rate of 90.783%, and 9,755,154 false positives. Surprisingly, this model performed substantially better on the test set than the validation set with a total model accuracy of 0.1668, a true positive rate of 96.92%, and 4,964,118 false positives. [report]

Investigating cheatgrass extent in the western US by Tamara Rudic

I use a time series of Landsat data spanning 2011-2020 and a comprehensive set of ground-truthed cheatgrass percent cover training data to investigate the efficacy of using machine learning models to detect cheatgrass presence and expansion in the Western US region. I select six different feature sets including both annual and seasonal spectral-temporal metrics and three different pairs of percent cover thresholds for defining cheatgrass absence versus presence to test two different baseline classifier models, random forest models, and the impacts of synthetic data augmentation using the borderlines SMOTE algorithm. I find that the application of synthetically augmented data is crucial for increasing random forest model performance relative to the baseline and using a threshold of 0%/20% to define cheatgrass absence versus presence (respectively) is most effective for increasing model performance. I find inconclusive results concerning the comparison of annual and seasonal metrics which suggest that annual metrics are preferred if high cheatgrass presence recall is the optimal performance metric, but seasonal metrics are preferred in high cheatgrass presence precision is the optimal performance metric. Finally, I suggest important future steps and limitations that need to be addressed in order to increase the performance of machine learning models on cheatgrass detection, including gathering ground-truthed data sets that are more representative of the study region at large and correcting for class imbalances between cheatgrass presence and absence classes. [report]

Detecting Deforestation with a Convolutional Neural Network Ensemble by Henry Talbott

Detecting deforestation from satellite imagery is a crucial problem both in terms of monitoring ecosystem health and predicting global and regional climate change. I trained multiple convolutional neural networks (CNNs) on year-2000 and year-2019 satellite images and accompanying deforestation data, with separate models trained on western Canada and the Amazon. I then created an ensemble model for each region by training a random forest classifier on the CNN outputs for each region, with the classification problem of determining is a given 2.4km by 2.4km patch of land was over 20% deforested. The Amazon ensemble model performed moderately well, with a test accuracy of 85% and a false positive rate of 10%. Both models had false negative rates lower than 10%, and outperformed baseline random forest and multilayer perceptron models. However, both ensemble models performed very poorly when tested on opposite regions of the ones they had been trained on, suggesting features learned during training were highly ecoregion-specific. Further research should focus on increasing the models' ability to generalize across distinct ecoregions. [report] [dataset] [data viz tool]

New approaches to understanding the structure of Alaska using machine learning by Isabella Gama

Seismology is a data rich science that is growing very fast. This growth is due to increasing technology capabilities and computational power. In the past five years, hundreds of seismometers were employed in Alaska, and today we have unprecedented amount of information about Alaskan subsurface. As scientists we need to find ways to keep up with this rapid technological growth by finding robust techniques that allow us to more accurately interpret these new data. The goal of this project is to use unsupervised learning techniques from machine learning to understand patterns of Earth’s subsurface with large amounts of seismic data recorded in Alaska. The data for this project have been processed to facilitate interpretation of results calculated here. I standardized the data, tested K-means and DBSCAN clustering algorithms, and implemented PCA and NMF dimensionality reduction methods. However, given the difficulties in interpretation of results from these unsupervised learning methods in terms of Earth’s characteristics, the main takeaway from this project it that it served as an opportunity to use all these method tests as exploration data analysis. [report]

Smart Models, Smart Meters by Matthew Park

Balancing the energy grid is crucial in preventing grid failure and blackouts. This balancing requires cooperation from consumers and suppliers of energy, demonstrating the importance of energy providers to understand consumer behavior. This project aims to predict the daily overall energy consumption of households based on a collection of weather metrics and a socio-demographic classification, referred to as the Acorn group. A collection of daily energy measurements for 5,567 London households from November 2011 to February 2014 were merged with daily numerical weather metrics and a series of engineered features. The data was used to train and evaluate a series of Regression models. The results demonstrate that it is possible to generate predictive models of daily energy consumption with accuracies of 0.8492 and 0.8812. The trained models generalized for the behavior of each Acorn classification in order to account for the variability in individual household behaviors. Similarly, the models revealed that the key feature in prediction was the Acorn classification. Thus, there is importance in understanding the relation between energy consumption and socio-demographic background, a feature itself which can be further broken down to distinguish and understand the behavior of an individual household. [report] [dataset]