Hierarchical classification with subsequent aggregation of heathland habitats using an intra-annual RapidEye time-series

International Journal of Applied Earth Observations and Geoinformation

Hannes Feilhauer¹  Marion Stellmes²  Björn Waske³  Kristin Fenske⁴  Michael Förster⁴ 

[1] Corresponding author at: Institute of Geographical Sciences, Freie Universität Berlin, Malteserstr. 74-100, 12249 Berlin, Germany.;Institute of Geography, FAU Erlangen-Nürnberg, Wetterkreuz 15, 91058 Erlangen, Germany;Geoinformation in Environmental Planning Lab, Technische Universität Berlin, Straße des 17. Juni 145, 10623 Berlin, Germany;Institute of Geographical Sciences, Freie Universität Berlin, Malteserstr. 74-100, 12249 Berlin, Germany;

关键词: Nature conservation;    Remote sensing;    Multispectral;    Multitemporal;    Grassland;    Synthetic map;   

DOI  :  

来源: DOAJ

【 摘 要 】

Mapping heathland habitats is generally challenging due to fine-scale habitats as well as spectral ambiguities between different classes. A multi-seasonal time-series of multispectral RapidEye data from several phenological stages was analysed towards the classification of different vegetation communities.A 3-level hierarchical dependent classification using Import Vector Machines was tested, based on the assumption that a probabilistic output per class would help the mapping. The first level of the hierarchical classification was related to the moisture gradient, which was derived from Ellenberg’s moisture indicative value. The second level aimed to separate plant alliances; the third level differentiated individual plant associations.For the final integration of the three classification levels, two approaches were implemented: (i) the F1-score and (ii) the maximum classification probability. The overall classification accuracies of both methods were found to be similar, around 0.7.Nevertheless, based on our expert knowledge we found the probabilistic approach to provide a more realistic picture and to be more practical compared to the result using the F1-score from the management point of view. In addition, the overall performance of the maximum probabilistic approach is better in the sense that the same accuracy of 0.7 was achieved with a differentiation of 33 classes instead of only 13 classes for the F1-score, meaning that the method is able to separate more spectral classes at a more detailed level providing the same accuracy.

【 授权许可】

Unknown