This paper examines strategies that can be used to determine the appropriate binarization of predictive land-cover maps to produce categorical land-cover maps, in this study used to separate peatland from non-peatland. Seven different strategies were applied to two predictive peatland maps, and the accuracy of the resultant binary land cover maps was evaluated. The main objective was to find the most effective approach to include as much peatland as possible, while simultaneously keeping the amount of noise (false positives) in the peatland map at a minimum. The best overall results were obtained with metrics related to correlation in the confusion matrix. Cohen’s Kappa and the F 1 score (defined as the harmonic mean of precision and recall) both reached their maximum at the same cutoff value, producing a land cover map with relatively high recall and limited amounts of noise in terms of false positive results. Maximizing the F 1 score does not necessarily produce the optimal result for all applications. The intended purpose of the map must also be considered when deciding whether it is more important to increase true positive results or minimize false positives. In this study, selecting the cutoff point by maximizing Cohen’s Kappa or the F 1 score proved to be the most effective overall strategy for dichotomizing the maps. Other strategies may be more appropriate when the distribution of the predictive scores is more balanced or there is a partisan preference for enhancing either user’s accuracy or producer’s accuracy. • We started with two remote sensing models showing probability for being peatland • Seven strategies for thresholding the probability scale to produce a binary peatland map were examined • Cohen’s Kappa and F 1 (the harmonic mean of precision and recall) reached their maximum at the same cutoff value • This cutoff value produced a map with relatively high recall and limited amounts of noise
Geir‐Harald Strand (2026) studied this question.