BOSTON UNIVERSITY GRADUATE SCHOOL OF ARTS AND SCIENCES Thesis BURNED AREA DETECTION ALGORITHM IN THE MODIS 500-METERS SURFACE REFLECTANCE DATA WITH SUPERVISED AND UNSUPERVISED NEURAL NETWORKS

Biomass burning is a key type of the Earth's land cover conversion. It destroys vegetation ecosystems, releases land surface carbon stocks to the atmosphere and alters biosphere-atmosphere interactions at regional and global scales. Accurate accounting of burned areas is critical for global carbon cycle modeling. It has unique value for timely management of natural resources. NASA's Earth Science Enterprise (ESE) provides required set of remote sensing measurements at the daily basis to allow nearly real time burned area monitoring at global scale. This study utilizes atmospherically corrected surface reflectances at 500-m spatial resolution recorded by the Moderate Resolution Imaging Spectroradiometer (MODIS) on board of the ESE Terra platform for burning area detection on the basis of the artificial intelligence approach. The artificial neural network systems, ARTMAP and Self-Organizing Map (SOM), were used for supervised and unsupervised extraction of burn features in the Earth's land cover. Training of ARTMAP system was conducted with MODIS land cover and fire products. Case studies include burn identification over Western regions of the US and Africa continent. We employ rough sets theory to provide a quantitative description of the accuracy of the feature detection system. This