Multi-temporal detection of grassland vegetation with RapidEye imagery and a spectral-temporal library

The detection of different vegetation classes especially in the context of biological diversity is one of the most frequently addressed topics in remote sensing studies. With the increasing availability of multi-temporal remote sensing information at high spatial resolution from Sensors such as RapidEye or future Sentinel-2, the phenological variability of vegetation classes can be exploited in more detail. At the same time, a larger number of repeated field spectral measurements are available for comparison with or classification of the image information. The presented study applies a spectral-temporal library (STL) for 21 different types of grassland species compositions to a set of 24 RapidEye scenes acquired in the vegetation periods between 2009 and 2011 for a test area located west of Berlin, Germany. Results of a Support Vector Machine (SVM) classification derived with information from the STL were compared with those of image information of training areas. The results show that a classification based on image spectra performs generally better (up to a Kappa of 0.82) while the application of the spectral library derived lower accuracies (below a Kappa of 0.5). However, the application of the STL performs very well for plant communities with a more homogeneous spectral behavior and limited anthropogenic influences, such as annual swaths.