Comparison of Sentinel-2 Multitemporal Approaches for Tree Species Mapping Within Natura 2000 Riparian Forest

Abstract

Highlights: What are the main findings? A comparison of three widely used tree species mapping approaches using multitemporal Sentinel-2 data revealed no statistically significant differences in classification accuracy; The approaches varied significantly in terms of the number of features required for model training and their potential for transferability. What is the implication of the main finding? Multitemporal classification models based on Sentinel-2 imagery can enhance the accuracy of riparian Natura 2000 habitat maps by identifying stands of non-native tree species and mitigating the overestimation of natural habitats; Riparian tree species maps can be translated into habitat classifications under the EU Habitats Directive; The transferability of tree species classification models to novel geographic regions and temporal contexts must be demonstrated to effectively support large-scale conservation planning, habitat quality assessment, and the evaluation of ecosystem services across Europe. Mapping forest tree species is vital for the habitat assessment, ecosystem services estimation, and implementation of European environmental policies such as the Habitats Directive. This study explores how repeated satellite observations over time, known as multitemporal data, can improve the mapping of tree species in riparian forests. Although many studies have shown that the use of multitemporal data improves tree species classification accuracies, there is a lack of research on how different multitemporal models perform compared to each other. We compared three multitemporal remote sensing approaches using Sentinel-2 imagery to map tree species within the Austrian riparian Natura 2000 site, Salzachauen. Seven tree species (five native and two non-native riparian species) were mapped using random forest models trained on a dataset of 444 validated tree samples. The three multitemporal approaches tested were: (i) multi-date image stacking, (ii) seasonal mean composites, and (iii) spectral–temporal metrics (STMs). The three approaches were compared to twenty single-date image classifications. The multitemporal models achieved 62 to 65% overall accuracy, while the median accuracy of single-date classification was 50% (SD = 6%). The seasonal model obtained the highest overall accuracy (65%), with F1 scores exceeding 73% for four individual species. However, differences among the three multitemporal approaches were not statistically significant. The mapping of native versus non-native riparian species achieved 92% accuracy. We evaluated misclassification patterns of individual species according to the two riparian forest habitats, 91E0* and 91F0, as defined in Annex I of the Habitats Directive. Most omission and commission errors occurred between species within the same habitat type. These findings underline the potential of translating tree species mapping to habitat-type classifications and the need to further explore the capabilities of satellite remote sensing to fill data gaps in Natura 2000 areas.