Master's thesis
Comparison of Sentinel-2 multitemporal approaches for tree species classification in Natura 2000: A case study of riparian forest mapping in Salzachauen protected area, Austria
by Yana Nikolova, 2025
Supervisor: Assoc. Prof. Dr. Hermann KLUG.
Department: Geoinformatics – Z_GIS
Attainment: Applied Geoinformatics Master's degree, at the Faculty of Digital and Analytical Sciences of the Paris-Lodron-University of Salzburg
The complete python code used for the analyses can be found on GitHub:
​
© Yana Nikolova
Abstract
​
Mapping the presence of less common but ecologically significant tree species is essential for ecosystem services estimation, environmental management, and informing policymakers on the coverage and conditions of distinct forest habitats. In this study, we mapped five native riparian tree species and two non-native species within the Salzachauen Natura 2000 area in Austria using multitemporal Sentinel-2 data and Random Forest classification. Twenty single-date and three multitemporal models were trained and tested to assess the best achievable accuracy for the area. We compared the three most commonly used multitemporal approaches for tree species remote sensing classification based on (i) multi-date stacked images, (ii) seasonal mean statistics, and (iii) annual spectral-temporal metrics (STMs). An equal ratio of reference data in pure and mixed pixels situations of the 20m resolution bands was used in the testing phase to ensure unbiased results and prevent artificially inflated accuracy. Overall, the multitemporal models achieved 62-65% overall accuracy, while the mean accuracy of single-date classification was 48%. The seasonal model obtained the highest overall accuracy, with an F1 score of four species reaching above 73%. Even though overall accuracies on single tree species classification did not show high results, the model achieved 92% accuracy in classifying native (autochthonous) versus non-native (allochthonous) riparian species. We evaluated over-/underestimation of single-species classifications according to the two riparian forest habitats 91E0* and 91F0, as defined in Annex I of the Habitat Directive. Classifications of single species were most often confused with species of the same habitat class. The essence and novelty of our work were the focus on less common but ecologically significant riparian tree species, as well as the comparison of three major multitemporal models. Our findings show 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.
© Yana Nikolova
Results
Figure 1: Tree species classification map obtained by RF classification of inter-seasonal stacked image. Overall accuracy of 65% where four species have reached above 73% F1 score accuracy. Those species are Picea abies, Populus balsamifera, Alnus incana and Salix alba.
Figure 2: Classification of non-native tree species Populus balsamifera and Picea abies, which intersects with Natura 2000 riparian forest habitats (91E0* and 91F0). Orthophotos are included for reference.
​
The multitemporal approach demonstrated a significant improvement in overall classification accuracy compared to single-date analyses, achieving above 73% accuracy for four individual species (Figure 1) and 92% overall accuracy for classifying native versus non-native species (Figure 2). The successful identification of non-native species within a Natura 2000 protected area emphasizes the practical applications of this methodology for environmental monitoring and management. By providing maps of native and non-native species distribution, this approach can support spatial planning for habitat restoration of riparian forests and inform policy decisions, particularly related to the Habitats Directive, by providing additional information on habitat distribution and conditions. To achieve this, more research focused on the less common but ecologically significant tree species is needed, accompanied by higher quantity and quality of reference data. The results also show high suitability for future implementation of class modelling of the 91E0* and 91F0 riparian forest habitats.
​
The findings highlight the importance of vegetation indices, especially those obtained during the early growing season (April and May), for effective tree species classification. However, the study also revealed challenges related to mixed pixels at 20x20m resolution, which influenced classification accuracy. Species with a higher ratio of pure pixels achieved better classification results, suggesting that future studies should consider both pure and mixed pixels in accuracy assessments to avoid bias and ensure a more comprehensive representation of the forest landscape.
​
Future work should also explore in more detail the scope of spectral temporal metrics for improving classification accuracies and transferability. Last but not least, translating such tree species classification models into habitat type maps, should be practically evaluated for the management and monitoring needs of data poor Natura 2000 areas.