In numerous applications of land-use/land-cover (LULC) classification, the classification rules are determined using a set of training data; thus, the results are inherently affected by uncertainty in the selection of those data. Few studies have assessed the accuracy of LULC classification with this consideration. In this article, we provide a general expression of various measures of classification accuracy with regard to the sample data set for classifier training and the sample data set for the evaluation of the classification results. We conducted stochastic simulations for LULC classification of a two-feature two-class case and a three-feature four-class case to show the uncertainties in the training sample and reference sample confusion matrices. A bootstrap simulation approach for establishing the 95% confidence interval of the classifier global accuracy was proposed and validated through rigorous stochastic simulation. Moreover, theoretical relationships among the producer accuracy, user accuracy, and overall accuracy were derived. The results demonstrate that the sample size of class-specific training data and the a priori probabilities of individual LULC classes must be jointly considered to ensure the correct determination of LULC classification accuracy.

<p><strong>Abstract.</strong> A majority of studies involving remote sensing LULC classification conducted classification accuracy assessment without consideration of the training data uncertainty. In this study we present new concepts of LULC classification accuracies, namely the training-sample-based global accuracy and the classifier global accuracy, and a general expression of different measures of classification accuracy in terms of the sample dataset for classifier training and the sample dataset for evaluation of classification results. Through stochastic simulation of a two-feature and two-class case, we demonstrate that the training-sample confusion matrix should replace the commonly adopted reference-sample confusion matrix for evaluation of LULC classification results. We then propose a bootstrap-simulation approach for establishing 95% confidence intervals of classifier global accuracies.</p>

No abstract

No abstract