全球变化知识资源中心

Updated and harmonized land cover (LC) data is essential for wildfire estimation in fire-prone areas as is the case in southern Europe. CORINE Land cover (CLC) and ESA Climate Change Initiative Land Cover (CCI-LC) maps have been analyzed and compared their performance in the estimation of wildfire occurrence in Europe at regional and local scales for the period 2010-2014. LC maps legends were harmonized and similarities and discrepancies were compared. Overall agreement between the two maps for the whole Europe was (similar to)75%. Forest and agriculture showed the largest agreement, while shrubland and grassland the lowest. Quantity and allocation disagreements were calculated including exchange and shift components (Pontius and Santacruz, 2014) which provided detailed information about the contribution of each class to the overall disagreement. Spatial discrepancies were found in areas where grassland and shrubland were the dominant classes as in United Kingdom or East Turkey. Land Use and Coverage Area frame Survey (LUCAS) was used as ground truth for validation purposes. The agreement with LUCAS was slightly higher for CCI-LC (59%) than for CLC (56%). Generalized Linear Models (GLM), based on presence-absence of wildfires, were used to estimate wildfire occurrence at 3 x 3 km grid cell resolution from both LC maps at the European scale. LC interfaces and climatic variables (temperature and precipitation) where used as explicative variables while fires from European Forest Fire Information System EFFIS (2010-2014 period) were used as response variable. Wildfire occurrence was also estimated with the two maps at local scale in a test region (Zamora, Spain) using a more precise location of the response variable (x, y fire ignition points). At the European scale models obtained within the two maps showed similar results. CCI-LC model sensitivity was 77.26%, specificity 25.89% and omission error 22.74% while CLC model sensitivity was 75.68%, specificity 29.99% and omission error 24.32%. However, CLC performed slightly better in terms of the percent correct classification (62%). In the test region the models achieved better results in terms of specificity (66.07% and 68.93% for CCI-LC and CLC models respectively) and percent correct classification ((similar to)68% for CLC model). At local scale CLC model performed better than CCI-LC model. Wildfire occurrence estimation was more accurate at local scale because of the differences in the spatial accuracy of the response variable used.