Researchers from the VSB-Technical University of Ostrava in the Czech Republic are working with teams from Argonne and Oak Ridge National Laboratories to identify wildfires in Alaska using satellite images and machine learning. Wildfires can have a significant impact on forest carbon balance, but are often difficult to quantify in remote regions such as boreal forests. To overcome this challenge, the researchers have developed a model to identify wildfires using data from remote sensing platforms and observatory networks.
The model employs a variety of tools, including Google Earth Engine API for obtaining satellite data and metadata, and the PERMON and PETSc frameworks for training classification models of a maximal-margin type. The PERMON software is designed to solve optimization problems of the quadratic programming type, supporting linear and bound constraints. It also supports computation on multiple GPUs, allowing for distributed settings.
Once the models have been calibrated, statistical inference is used to transform an uncalibrated output into a probability of class membership. This is achieved using the Platt scaling approach based on cross-entropy minimization. Marek Pecha, PhD candidate in Applied Mathematics at VSB-Technical University of Ostrava, and member of the PERMON team, is leading the project. He focuses on extending the functionality of the PERMON toolbox for training machine learning models, and leads a small team at the Institute of Geonics at Czech Academy of Sciences.
Their work is supported by the Strategy AV21 project, where they are developing machine learning approaches for processing and analyzing seismic events. The researchers are also discussing adaptations of the solvers for optimization problems arising from these models, with the ultimate aim of improving the approach using deep learning techniques.
