Grassland Identification Using Multi-Temporal RapidEye Image Series
Weichelt, Horst; Zillmann, Erik
The incorporation of multi-temporal data into image classification can help to overcome the problem of discriminating land cover classes that have similar spectral signatures but different phenological dynamics, such as the case of cropland and grassland. For the Global Monitoring of Environment and Security (GMES) land monitoring services a multi-temporal classification approach used for pan-European grassland discrimination has been developed and carried out successfully for the European Environmental Agency (EEA). The methodology is based on a multi-temporal analysis of multi-scale image series of three different multi-spectral sensors LISS III, AWiFS and RapidEye. The RapidEye satellite constellation represents a unique potential of multi-temporal acquisition of high resolution image data, therefore, offering completely new ways of detailed multi-temporal analysis. The full utilization of this information for land-cover classification is yet to be brought forward and new approaches need to be developed or existing methods adapted. The objective of the presented study is to demonstrate the applicability of the developed grassland classification approach using solely multi-temporal RapidEye image series. TOA calibrated reflectance images of 2011 were used to achieve consistency throughout the entire time series and to retrieve a set of biophysical parameters characterizing the spatio-temporal behavior of vegetation cover and growth status. In order to fully utilize the information of the multi-temporal data set, various statistical features of each biophysical parameter were calculated and used as the main basis for subsequent classification. A supervised object-based image classification using different object features was adopted to classify grasslands. A Classification and Regression Tree (CART) classifier was used to analyze the data and assign the terrain classes. The CART approach is well suited for non-normally distributed data, as it happens when each land-cover class is represented by more than one cluster in the spectral feature space. The results presented correspond to an area of 2500 km2 in the State of Brandenburg / Germany. The landscape is mainly characterized by a mixture of forest and grassland. The accuracy of the grassland classification was assessed by using the administrative field block cadaster (MIL Brandenburg, Feb.2012). The grassland classification of the test area reached an overall accuracy of about 90%.