Abstract: Wildfires belong to important sources of disturbance and can significantly affect forest carbon balance. It can be difficult to accurately quantify the effects of wildfire in places such as boreal forests that are far from human habitation and infrastructure. Data from remote sensing platforms and observatory networks can be used for this task. However, these data sets can be vast, and analyzing them can require powerful computing resources and tools designed to fully utilize them.
In this talk, we will present an ongoing cooperation among PERMON team from VSB-Technical University of Ostrava and Institute of Geonics Czech Academy of Sciences (Czech Republic), Richard Mills from Argonne National Laboratory, and Zachary Langford from Oak Ridge National Laboratory. We will go through the current stage of development and the results of our approach to wildfire identification. Our software solution employs Google Earth Engine API for getting multi-spectral satellite data and metadata, i.e. labels related to wildfires used for training semantic segmentation models. We train these models in distributed settings supporting computation on multiple GPUs using our in-house software written on the top of the PETSc framework (https://petsc.org/). This software is called PERMON (http://permon.vsb.cz). It is designed for solving the optimization problems of the quadratic programming (QP) type supporting linear and bound constraints. Based on the PERMON and PETSc frameworks, we implement a tool called PermonSVM for training classification models of a maximal-margin type, namely Support Vector Machines (SVM). A training procedure uses our efficient QP solvers developed in Prof. Dostál’s group, which are implemented in PERMON, during the training phase of these models, or alternatively, compatible solvers from TAO (The Toolkit for Advanced Optimization) could also be employed. We will discuss adaptations of these solvers for optimization problems arising from these models, e.g. designing suitable stopping criteria.
Afterwards, the models could be calibrated. Calibration of a classification model refers to a special type of statistical inference that transforms an uncalibrated output (distances of samples from separating hyperplane) to a probability of class membership. We will introduce the Platt scaling approach based on cross-entropy minimization, where a posterior probability is approximated by means of logistic regression. It leads to an unconstrained optimization we can solvable by using the Newton type methods implemented in TAO effectively.
The introduction of techniques for data transformation and their impacts on model performance and ideas on how this approach could be improved using deep learning will also be discussed in a presentation.
Bio: Marek Pecha is PhD candidate of Applied Mathematics at VSB-Technical University of Ostrava, Czech Republic, under supervision of David Horák and co-supervision of Richard Mills (MCS). Pecha is a member of the PERMON team, where he focuses on extending the functionality of the PERMON toolbox for training machine learning models. He also leads a small team at the Institute of Geonics at Czech Academy of Sciences. Their work is supported by the Strategy AV21 project (https://strategie.avcr.cz/en), where they work on processing and analyzing seismic events using machine learning approaches.