Stenzel, Stefanie Anne: Multiseasonal Remote Sensing of Vegetation with One-Class Classification – Possibilities and Limitations in Detecting Habitats of Nature Conservation Value. - Bonn, 2017. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc: https://nbn-resolving.org/urn:nbn:de:hbz:5n-47585
@phdthesis{handle:20.500.11811/7197,
urn: https://nbn-resolving.org/urn:nbn:de:hbz:5n-47585,
author = {{Stefanie Anne Stenzel}},
title = {Multiseasonal Remote Sensing of Vegetation with One-Class Classification – Possibilities and Limitations in Detecting Habitats of Nature Conservation Value},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2017,
month = jul,

note = {Mapping of habitats relevant for nature conservation often involves the identification of patches of target habitats in a complex mosaic of vegetation types extraneous for conservation planning. In field surveys, this is often a time-consuming and work-intensive task. Limiting the necessary ground reference to a small sample of target habitats and combining it with area-wide remote sensing data could greatly reduce and therefore support the field mapping effort. Conventional supervised classification methods need to be trained with a representative set of samples covering an exhaustive set of all classes. Acquiring such data is work intensive and hence inefficient in cases where only one or few classes are of interest. The usage of one-class classifiers (OCC) seems to be more suitable for this task – but has up until now neither been tested nor applied for large scale mapping and monitoring in programs such as those requested for the Natura 2000 European Habitat Directive or the High Nature Value (HNV) farmland Indicator. It is important to uncover the possibilities and mark the obstacles of this new approach since the usage of remote sensing for conservation purposes is currently controversially discussed in the ecology community as well as in the remote sensing community. Thus, the focal and innovative point of this thesis is to explore possibilities and limitations in the application of one-class classifiers for detecting habitats of nature conservation value with the help of multi-seasonal remote sensing and limited field data.
The first study ascertains the usage of an OCC is suitable for mapping Natura 2000 habitat types. Applying the Maxent algorithm in combination with a low number of ground reference points of four habitat types and easily available multi-seasonal satellite imagery resulted in a combined habitat map with reasonable accuracy. There is potential in one-class classification for detecting rare habitats – however, differentiating habitats with very similar species composition remains challenging.
Motivated by these positive results, the topic of the second study of this thesis is whether low and HNV grasslands can be differentiated with remotely-sensed reflectance data, field data and one-class classification. This approach could supplement existing field survey-based monitoring approaches such as for the HNV farmland Indicator. Three one-class classifiers together with multi-seasonal, multispectral remote sensing data in combination with sparse field data were analysed for their usage A) to classify HNV grassland against other areas and B) to differentiate between three quality classes of HNV grassland according to the current German HNV monitoring approach. Results indicated reasonable performances of the OCC to identify HNV grassland areas, but clearly showed that a separation into several HNV quality classes is not possible.
In the third study the robustness and weak spots of an OCC were tested considering the effect of landscape composition and sample size on accuracy measurements. For this purpose artificial landscapes were generated to avoid the common problem of case-studies which usually can only make locally valid statements on the suitability of a tested approach. Whereas results concerning target sample size and the amount of similar classes in the background confirm conclusions of earlier studies from the field of species distribution modelling, results for background sample size and prevalence of target class give new insights and a basis for further studies and discussions.
In conclusion the utilisation of an OCC together with reflectance and sparse field data for addressing rare vegetation types of conservation interest proves to be useful and has to be recommended for further research.},

url = {https://hdl.handle.net/20.500.11811/7197}
}

The following license files are associated with this item:

InCopyright