Using logit models to classify land cover and land-cover change from Landsat Thematic Mapper

In this paper, we use logit models to classify data from Landsat Thematic Mapper (TM) among 23 land-cover and land-cover change classes. The logit model is a simple statistical technique that is designed to analyse categorical data. Diagnostic statistics indicate that the logit model can classify remotely sensed data in a statistically significant fashion. User accuracies for individual land-cover classes range between 50 and 92%, with an overall accuracy of 79%. To assess these accuracies, we compare them to those generated by a Bayesian maximum likelihood classifier. While the overall accuracies are similar, the accuracies for individual land-cover categories differ. These differences may be associated with the size of the training data for each land-cover class. There is some evidence that the logit models generate higher accuracies for land-cover categories for which relatively few training pixels are available. Finally, a comparison of classification results using a 12-band composite of the six reflective TM bands and their change vectors versus a six-band composite of the three Tasselled Cap bands and their change vectors indicates that the latter reduces classification accuracies.