Research

Clustering Strategies For XAI with Correlated, High-Dimensional Geoscience AI Models

With the rapid deployment of high-resolution sensors and models, geospatial data is captured and generated at an extremely high rate. By automatically extracting information from large data volumes, machine learning is increasingly used to turn massive geospatial data into geoscience insights. There is widespread use of high-dimensional raster data from which complex machine learning algorithms learn to recognize spatial and spatial-temporal patterns. These models may be used for critical decision-making or as a way to aid scientific discovery. Complex models are able to learn high non-linear relationships to achieve high performance, but there is a concern that their complexity makes it very difficult for users to determine how the model reached its decision. This has motivated the widespread adoption of explainable artificial intelligence (XAI) techniques that probe the model in various ways to explain how it works.

XAI methods are highly sensitive to correlations among input predictors. Proposed mitigations involve grouping correlated predictors and applying XAI to groups instead of individuals. However, there are major challenges to grouping the grid cells of geoscience rasters based on their correlation. These datasets are commonly high-dimensional with substantial autocorrelation. Conventional techniques for grouping correlated tabular features are rarely applicable.

The purpose of this research is to develop strategies for using data-driven clustering techniques to group raster data to improve the accuracy of XAI results. First, we describe the limitations of current approaches and identify XAI challenges related to raster-based geoscience models. We then develop a set of benchmarks so that we can quantitatively assess XAI methods in a variety of complex scenarios. Finally, we propose and evaluate methodologies for applying clustering and XAI techniques. These include a hierarchical clustering approach to automatically investigate multiple scales of patterns in high-dimensional data.

Publications

Presentations

Energy Efficient Path Planning for Autonomous Surface Vessels

Typical robot path planning is based on optimizing the path to achieve that with the shortest distance. For a robot boat, however, the influence of the environment (water currents, waves, etc) may be of much greater concern. In this research, we explore multiple approaches to path planning for an autonomous surface vessel when taking into consideration ocean current forecasts. These approaches include metaheuristic algorithms and game-theoretic techniques.

Publications

Repositories

Presentations

Machine Learning Pipeline for Detection and Prediction of PyroCbs

Under favorable atmospheric conditions, intense heat from a large and hot wildfire can generate deep, smoke-infused storms resembling conventional thunderstorms that are known as pyrocumulonimbus (pyroCb). PyroCbs are capable of releasing a large quantity of smoke particles into the lower stratosphere, often above the tropopause as well as above jet aircraft cruising altitudes by several kilometers. PyroCbs can be accompanied by strong and erratic inflow, potentially dangerous downbursts, and lightning strikes. These extreme events can increase fire spread rates and intensity, cause sudden changes in fire spread directions, and ignite additional fires. These storms are especially dangerous for fire fighters and others involved in disaster response. In this research, we have developed a machine learning system for understanding and detecting the atmospheric potential of wildfires in producing proCb as Wildfire–driven Thunderstorms. This is challenging because pyroCbs are extreme events: there are ~500 incidents during the 2013 – 2021 period. Machine learning typically struggles to learn from relatively few examples and highly imbalanced datasets. Our pipeline involves fusing input data sources to assemble a training dataset, applying feature selection and data balancing techniques, and training with several machine learning algorithms. Finally, we perform initial eXplainable AI (XAI) experiments to analyze learned model strategies.

Presentations