7 th EARSeL Workshop on Imaging Spectroscopy 1 Edinburgh, Scotland, 11 th – 13 th April, 2011 AUTOMATIC FOREST AREA EXTRACTION FROM IMAGING SPECTROSCOPY DATA USING AN EXTENDED NDVI Fabian Faßnacht 1 , Holger Weinacker 2 and Barbara Koch 3 1. University of Freiburg, Department FELIS, Freiburg, Germany; fabian.fassnacht@felis.uni- freiburg.de 2. University of Freiburg, Department FELIS, Freiburg, Germany; holger.weinacker@felis.uni- freiburg.de 3. University of Freiburg, Department FELIS, Freiburg, Germany; barbara.koch@felis.uni- freiburg.de ABSTRACT A method is presented to automatically extract tree-covered areas from airborne and simulated spaceborne imaging spectroscopy data. The method is based on the extended Normalized Differ- ence Vegetation Index (NDVI). The function of the index is to continuously decrease the index-values of all land surface classes in relation to the reflection values of tree-covered areas resulting in an index image with tree covered areas having highest values. Besides the typical wavelengths of the NDVI in the visual red and the near infrared section of the spectrum (660nm & 760 nm) two additional channels in the near infra- red (810nm & 2450nm) were selected to boost the NDVI. Mean reflectance values of different land use surface classes were used to scale and weight the reflectance values within the selected channels resulting in values similar to the outcome of the NDVI (values from -1 to 1). Based on the calculated index-image and an additional variance-image derived from the index- image a binary mask was created using a threshold value for each of the input images. The method was applied to 19 HyMap-Scenes with ground sampling distances (GSD) varying be- tween 4m and 8m. Additionally the index was tested on a simulated EnMAP scene with 30m GSD. After calibrating the index to the sensor by including the mean reflectance values of only a few training areas collected in two HyMap-scenes the method delivered promising results over all 19 HyMap scenes collected in various regions of Germany. For usage with the simulated EnMAP scene the threshold value of the variance image had to be adapted. Forest / non-forest classifica- tion accuracies of four statistically evaluated scenes reached from 92.0% to 96.6%. Almost all mis- classified samples could be assigned to one of the following three classes: 1. Areas influenced by BRDF effects at the border of the image or in pronounced terrain situations 2. Swampy areas 3. Agricultural areas (esp. maize). INTRODUCTION Forests play a major role in the world’s terrestrial ecosystems as well as in worldwide climatic processes. Forest ecosystems provide a wide range of material and nonmaterial benefits such as wood, fibre, game, pharmaceutical usable herbs, oxygen production, carbon absorption, habitat, erosion protection and many more. To protect and sustain forests as an efficient source of the aforementioned benefits, sustainable management practices and the reduction of deforestation and forest degradation has to be in the focus of forestry practitioners and researchers. Worldwide activities such as the United Nations Collaborative Programme on Reducing Emissions from De- forestation and Forest Degradation in Developing Countries (REDD) reflect the worldwide aware- ness of this important issue. Remote Sensing with its ability to collect continuous data over wide areas is an important tool in assessing the actual state of worldwide forests and can be used as an efficient source of informa- tion on local, regional and global scale.