Multiple Instance Learning for Burned Area Mapping Using Multi-Temporal Reflectance Data

Guruprasad Nayak, University of Minnesota

Mapping burned area on a global scale typically requires the use of a weak signal like Active Fire for training the burned scar classification model. Since these weak signals typically are inaccurate with respect to temporal and spatial pinpointing of the event occurrence, the use of Multiple instance learning paradigm to model the occurrence of the event in a wider spatio-temporal window is demonstrably beneficial than using the exact date of the weak signal. In this work, we demonstrate the use of MIL algorithm to model the temporal uncertainty of the weak signal. We further propose an noise-robust extension to the MIL paradigm for learning on sequence data.

Link to Presentation: CI2016 Nayak