TEMPORAL CONTEXT IN FLORISTIC CLASSIFICATION

Multi-temporal remote sensing data present a number of significant pro blems for the statistical and spatial competence of a classifier. Idea lly, a classifier of multi-temporal data should be temporally invarian t. It must have the capacity to account for the variations in season, growth cycle, radiometric, and atmospheric conditions at any point in time when classifying the land cover. This paper tests two methods of creating a temporally invariant classifier based on the pattern recogn ition capabilities of a neural network. A suite of twelve multi-tempor al datasets spread over 5 yr along with a comprehensive mix of environ mental variables are fused into floristic classification images by the neural network. Uncertainties in the classifications are addressed ex plicitly with a confidence mask generated from the fuzzy membership va lue's output by the neural network. These confidence masks are used to produce constrained classification images. The overall accuracy perce ntage achieved from a study site containing highly disturbed undulatin g terrain averages 60%. The first method of training, sequential learn ing of temporal context, is tested by an examination of the step-by-st ep evolution of the sequential training process. This reveals that the sequential classifier may not have learned about time, because time w as constant during each network training session. It also suggests tha t there are optimal times during the annual cycle to train the classif ier for particular floristic classes. The second method of training th e classifier is randomised exposure to the entire temporal training su ite. Time was now a fluctuating input variable during the network trai ning process. This method produced the best spatially accurate results . The performance of this classifier as a temporally invariant classif ier is tested amongst four multi-temporal datasets with encouraging re sults. The classifier consistently achieved an overall accuracy percen tage of 60%. The pairwise predicted overall accuracy percentage averag ed 80%. The randomised trained neural network seems robust against the variations of season, radiometric, and atmospheric conditions, and sh ows great promise as a temporally invariant classifier. Copyright (C) 1996 Elsevier Science Ltd.