Evaluating Rule Set Transferability for Extracting Forest Resources from RapidEye Data
Kindu, Mengistie1; Elatawneh, Alata2; Corti, Nicolas3; Rappl, Adelheid4; Felbermeier, Bernhard5; Acevedo Cabra, Ricardo6; Schneider, Thomas5; Knoke, Thomas5
1TU München, ETHIOPIA; 2TU München, PALESTINIAN TERRITORIES; 3TU München, ECUADOR; 4Bavarian State Institute of Forestry, GERMANY; 5TU München, GERMANY; 6TU München, COLOMBIA

Knowledge of forest resource cover is a key factor as an input for a variety of applications including the current global climate change issue and initiatives like REDD+. This study analyzed the transferability of developed rule sets on classifying forest resources using RapidEye data from German (Oberammergau/Freising/Traunstein test sites) to other environmental settings; Ethiopia (Munessa), Ecuador (San Francisco) and China (Shangnan). Similar preprocessing steps were applied in each image for all test sites (geometry, atmosphere, topography). The rule sets were developed using object-based classification approach (OBIA) with a result of an overall accuracy 93.5% for forest and non-forest classification at level 1 and 89.5% for forest type classifications (broadleaved/ coniferous dominated/mixed, plantation/managed/close to nature forest) at level 2. Comparisons of transferability were conducted using accuracies of the classified images. The result reveals that forest and non-forest classes at level 1 classified with an overall accuracies of 91.7%, 86.1% and 71.2%, for Ethiopian, Ecuadorian and Chinese test sites, respectively. This means the developed rule sets can be directly transferred with an acceptable accuracy level for Ethiopian and Ecuadorian test sites. The low agreement for the Chinese test site is connected to the disturbed nature of the forest under investigation. The results of forest type classification at level 2 also show the need of further refinement of the already developed rule sets when transferred to a geographically different area. Adding more rules or adapting to each of the environmental settings is recommended for higher accuracy as required for a standardized evaluation of the resource forest in the global context. An easily to adapt version of the rule set was developed, allowing to adjust selected settings by local forest experts even without remote sensing practice.