This paper presents an ARTMAP neural network approach for burn detection in Moderate Resolution Imaging Spectroradiometer (MODIS) data using two methods: discrete and continuous classifications. The study area covers the states of Idaho and Montana in the United States, where extensive fire events took place during the months of July and August in the year 2000. The proposed approach differs from commonly used change detection schemes by utilizing a single surface reflectance image instead of time series of satellite data. Burn detection in this study was accomplished by the classification of land into four classes: burns, woody vegetation, herbaceous vegetation, and barren. We performed the discrete classification of coarse (500-m MODIS data) and high-resolution (30-m Enhanced Thematic Mapper (ETM+) data) surface reflectance data with an ARTMAP classifier to evaluate the impact of a land cover mixture on burn detection. The analysis of classification results reveals commission and omission errors in the evaluation of burn area extent at a coarse-resolution scale. To account for land cover heterogeneity, we utilized the continuous classification of coarse-resolution data with an ARTMAP mixture model. A training data set on the land cover mixture at 500-m scale of MODIS data was assembled from the aggregated 30-m ETM+ classification. The ARTMAP mixture model was trained with MODIS surface reflectance data and land cover mixture information to generate a continuous classification of burns (expressed in percentage of burns per pixel). Data fusion of coarse- and high-resolution satellite data in this study resulted in a more natural and accurate mapping of burns as mixtures with other land cover types.