Global Monitoring of Tropical Forest Fires Using A New Predictive Modeling Approach for Rare Classes

Abstract:

The traditional classification approaches use labeled training data to select the best classification model from a family of models. Since collecting labeled samples is often tedious and sometimes even infeasible, recent research in machine learning has focused on developing algorithms to train classification models in scarcity of labeled training samples. In contrast, the focus of our research is to address problem settings where acquiring even a small number of expert-annotated labeled samples for supervision is infeasible. I will present RAPT, a new predictive modeling framework for identifying rare classes when there is a complete absence of labeled data. The RAPT framework is designed to use imperfectly annotated training data to learn classification models in the absence of expert-annotated training samples. Our results show that, under some reasonable assumptions, the classifiers trained from imperfectly labeled training data using the RAPT approach have performance comparable to the classification models trained using expert-annotated training data. This capability of learning from imperfect supervision is advantageous in a wide range of applications where the target class of interest is relatively rare and obtaining a precise labeling of even a small number of training samples is infeasible. I will also present the application of the RAPT framework for creating historical maps of forest fires from satellite data for the tropical forests. This new forest fire product identifies approximately 1 million sq. km. of burned areas in the tropical forests in South America and South-east Asia during years 2001-2014, which is more than three times the total burned area reported by the state-of-art NASA products. We show validation of these results using burn-scars visible in satellite images, including high resolution Landsat images, to confirm the veracity of the previously unreported forest fires.