Fuzzy ARTMAP classification of global land cover from AVHRR data set

Phenological differences among broadly defined vegetation types can be a basis for global scale classification at a very coarse spatial scale. Using the annual sequence of composited Normalized Vegetation Index (NDVI) values in an AVHRR data set, DeFries and Townshend (1994) distinguished eleven global cover types and classified global land cover with a maximum likelihood classifier. The present research presents a neural network architecture called fuzzy ARTMAP to classify the same data set. Classification results are analyzed and compared with those obtained from DeFries and Townshend (1994). First, classification accuracy on the training data set is 100% compared with the corresponding accuracy of 86.8%. Second, when fuzzy ARTMAP is trained using 80% of data set and tested on the remaining (unseen) 20% data set, classification accuracy is more than 80%. Third, classification results vary with and without latitude as an input variable similar to those of DeFries and Townshend