Muchoney, DM, Strahler, AH (2002). Pixel- and site-based calibration and validation methods for evaluating supervised classification of remotely sensed data. REMOTE SENSING OF ENVIRONMENT, 81(3-Feb), 290-299.
The characteristics of calibration and validation data, especially sample size, distribution, thematic labeling, and representativeness, are important to supervised classification algorithms, as are their use to ascribe accuracy statements to supervised classification results. While random and stratified random sampling of calibration and validation into calibration (training) and validation (testing) subsets has often been based on pixel sampling, the resulting accuracy statements may be overly optimistic and biased due to spatial autocorrelation. This is especially true for decision tree and neural network algorithms that tend to learn examples well even when just a single pixel from a site is presented in the learning phase. Therefore, polygon- or site-based sampling (using groups of contiguous pixels) for calibration and validation may provide better estimates of prediction accuracies. This paper presents the results of pixel- versus polygon-based calibration and validation of a supervised Gaussian ARTMAP neural network algorithm. Both techniques were applied to (1) a multitemporal AVHRR vegetation index data set and (2) a multitemporal data set of NDVI, temperature, and additional temporal metric and physical data to map vegetation and land cover (VLC) based on a regional network of sites in Central America. Our results indicate that although pixel-based accuracy assessment because of spatial autocorrelation may overstate accuracy, polygon-based assessments are also problematic. The Gaussian ARTMAP algorithm was used to provide insight into the problems of site heterogeneity, per-class accuracy, sampling rate, sampling representativeness, and the generalization properties of sites and classes, as related to pixel- and polygon-based sampling. (C) 2002 Elsevier Science Inc. All rights reserved.