Let your maps be fuzzy!—Class probabilities and floristic gradients as alternatives to crisp mapping for remote sensing of vegetation

Mapping vegetation as hard classes based on remote sensing data is a frequently applied approach, even though this crisp, categorical representation is not in line with nature's fuzziness. Gradual transitions in plant species composition in ecotones and faint compositional differences across different patches are thus poorly described in the resulting maps. Several concepts promise to provide better vegetation maps. These include (1) fuzzy classification (a.k.a. soft classification) that takes the probability of an image pixel's class membership into account and (2) gradient mapping based on ordination, which describes plant species composition as a floristic continuum and avoids a categorical description of vegetation patterns. A systematic and comprehensive comparison of these approaches is missing to date. This paper hence gives an overview of the state of the art in fuzzy classification and gradient mapping and compares the approaches in a case study. The advantages and disadvantages of the approaches are discussed and their performance is compared to hard classification (a.k.a. crisp or boolean classification). Gradient mapping best conserves the information in the original data and does not require an a priori categorization. Fuzzy classification comes close in terms of information loss and likewise preserves the continuous nature of vegetation, however, still relying on a priori classification. The need for a priori classification may be a disadvantage or, in other cases, an advantage because it allows using categorical input data instead of the detailed vegetation records required for ordination. Both approaches support spatially explicit accuracy analyses, which further improves the usefulness of the output maps. Gradient mapping and fuzzy classification offer various advantages over hard classification, can always be transformed into a crisp map and are generally applicable to various data structures. We thus recommend the use of these approaches over hard classification for applications in ecological research.