Forest stratification in Fiji using very high resolution imagery

Summary

This report is a comprehensive summary of work done on the topic Object-Oriented Forest Stratification for REDD-readiness in Fiji for the Dogotuki forest region in Vanua Levu, Fiji. The study of forest stratification in Fiji using very high resolution satellite imagery was funded through the project “Climate Protection through Forest Conservation in Pacific Island Countries” on behalf of of the Deutsche Gesellschaft für Internationale Zussamenarbeit (GIZ) with funding from the German Federal Environment Ministry and performed in cooperation with Fiji Forest Department Management Services Division (MSD) and the Applied Geoscience and Technology Division of the Secretariat of the Pacific Community (SOPAC) in Suva, Fiji. For REDD-readiness the government of Fiji needs to develop a national carbon monitoring system. For this monitoring system it is necessary to develop a framework and method to stratify the Fijian forest according to chosen FAO forest definitions. In this research a method is proposed to go from a Worldview-2 satellite image, through a forest canopy map, to a forest density map. A definition dependent, object-based method is evaluated for the delineation of the areas of closed, open and non-forest. This method resulted in a density map that indicates a forest canopy cover less than 10% different from the actual canopy cover, an improvement compared to the results of other pixel-based methods. It has also been explored whether mangrove and pine can be distinguished spectrally, but this was not validated due to a lack of ground reference data. The parameters used for segmentation and classification of the canopy map proved to be usable independent of geographic aspects, showing that methods developed in one area can be applied on different areas without a significant loss in accuracy. The red band was used for segmentation and the NDVI, hue, saturation and maximum difference in brightness were shown to be the features most influential for the classification of the forest canopy. The automatic classification with a Bayesian classifier had an average overlap of 84% with visual interpretation and 76% with ground reference data.