Detection of trees outside forest (TOF) using digital aerial images – a cross-country approach

The need for up-to-date and accurate information on trees outside forest (TOF) has increased in recent years since they play a key role in estimating biomass, regarding degradation of landscape, climate change and biodiversity issues. Assessing TOF using national forest inventories (NFIs) databases is often not feasible since most NFIs are established and predestinated for forest. Using remote sensing methods might open new perspectives for detecting non-forest wood resources, such as TOF, other wooded land (OWL), and other land with tree cover (OLWTC) as defined by the Food and Agriculture Organization of the United Nations (FAO). Using remote sensing data sets also facilitate cross-border applications whereas the combined use of NFI data from different countries is challenging. In the present study that was part of the COST (Cooperation of science and technology) Action FP1001 ‘Usewood’ a robust method was developed that enables a semi-automated extraction of TOF based on airborne stereo imagery from Austria (Ultracam XP), Germany (Ultracam XP) and Switzerland (ADS80) in addition to terrain models from airborne laser scanning (ALS) and the corresponding forest definitions including FAO. As basic datasets wooded area masks were extracted from an NDVI image in combination with a canopy height model (CHM) using the height criteria of the four different forest definitions. This CHM was calculated from the individual datasets – digital surface and terrain models (DSMs, DTMs) - of the participating countries. The DSMs with a resolution of 1 x 1m were created with the help of Image Matching techniques using different software packages. Four different forest definitions (NFI Austria, NFI Germany, NFI Switzerland and FAO) were applied to the wooded area masks using the criteria ‘minimum crown coverage’, ‘minimum width’ and ‘minimum area’. The resulting maps included forest, OWL and OLWTC which fulfilled the respective definitions, remaining trees were regarded as TOF . Land use criteria had to be taken into account in order to separate forest from OWL, OLWTC and TOF. Therefore auxiliary data obtained from topographic maps of the three countries were used. All maps were validated using an independent reference data set which consisted of visual image interpretation at 1100 points in a regular grid. Accuracies between 70% to 95% were achieved for the different wood resource classes. It turned out that differences between the four resulting datasets were highly correlated with the forest definitions used. This study revealed that the cross-border approaches are robust and feasible with a high degree of automation. Although three different sets of aerial images were used and three different software packages for Image Matching were applied on the data almost no border effects could be detected. Whereas the extraction of wooded area is accurate, semi-automated and independent from the four used forest definitions, TOF is less accurate, since it depends on the used forest definitions and the land use information of topographic maps. The main disadvantage of Remote Sensing techniques is that land use according to forest definitions cannot be obtained. Therefore auxiliary topographic data sets were used. However, those data sets were different in acquisition date and level of detail compared to the used aerial images. The proposed method to extract wooded area using high resolution remote sensing data and to classify into well defined classes as forest, OWL, OLWTC and TOF might help for a harmonized reporting and cross-national comparison.