A rule-based system for burned area mapping in temperate and tropical regions using NOAA/AVHRR imagery

The feasibility of deriving a single classifier capable of mapping burned areas in Iberia and central Africa, using NOAA/AVHRR satellite imagery was investigated. A supervised classification approach based on the Classification and Regression Trees (CART) algorithm was used to classify a single date image from Africa and a multi-temporal composite from Iberia into three classes: burned, unburned, and cloud. A classification tree with 22 terminal nodes constructed was constructed with CART, using albedo, GEMI, and channel 4 brightness temperature as independent variables. The accuracy of this classifier was assessed on a set of independent data and found to be higher than 98% for each class. All burned area pixels in the test data set were correctly classified. Visual comparison with high resolution data in the case of Iberia, and with active fire data in the case of Africa, confirm this good performance. Advantageous characteristics of rule induction approaches for the classification of satellite imagery are discussed from the perspective of deriving a system for global burned area mapping.