Forest mapping with a generalized classifier and Landsat TM data

Monitoring landcover and landcover change at regional and global scales oft en requires Landsat data to identify and map landscape features and pattern s with sufficient detail. Analytical methods based on image-by-image interp retation are too time-consuming and labor-intensive for studies of large ar eas to be undertaken with any degree of frequency. One potential solution i s to develop algorithms or classifiers that can be generalized beyond the a rena of the initial training to new images from different spatial, temporal or sensor domains. Building upon earlier success with a generalized classi fier to monitor forest change, we now address the question of generalizatio n for classifications of stable landcovers. We evaluate the ability of a su pervised neural network, Fuzzy ARTMAP, to identify conifer forest across ti me and space with Landsat Thematic Mapper (TM) images for a region in north west Oregon. We also assess the effects of atmospheric corrections on gener alized classification accuracies. Using midsummer images atmospherically co rrected with a simple dark-object-subtraction (DOS) method, there is no sta tistically significant loss of accuracy as the classification is extended f rom the initial training image to other images from the same scene (path an d row): temporal generalization is successful. Extending the classifier acr oss space and time to nearby scenes results in a mean decline of 8-13% accu racy depending on the atmospheric correction used. Obvious sources of error , such as seasonality solar angle variation, and complexity of landcover id entification, do not explain the decline in error. Additionally, the patter ns in generalization accuracies are complex, and the relationship between p airs of training and testing images is not necessarily reciprocal, i.e., go od training data are not necessarily good testing data. Simple DOS atmosphe ric corrections produce classifications with comparable accuracies as class ifications from the more complex radiative transfer corrections. These find ings are based on over 200 classifications. A high degree of variability in the classification accuracies underscores the importance of extensive, in- depth analysis of remote sensing techniques and applications, and highlight s the potential problem for misleading results based on just a few tests. G eneralization is well suited for multitemporal classifications of one Lands at scene. Using simple DOS and midsummer images, generalization offers the opportunity for frequent landcover mapping of a Landsat scene without havin g to retrain the classifier for each time period of interest. However, at t his point, the utility of regional landcover mapping with a generalized cla ssifier remains limited. (C) 2001 Elsevier Science Inc. Alt rights reserved .