Pan-European Forest Maps Derived from Optical Satellite Imagery

Kempeneers, P. 1, Sedano, F. 2, Pekkarinen, A. 3, Seebach, L. 4, Strobl, P. 5, San-Miguel-Ayanz, J. 5
1 VITO NV, Boeretang 200, BE-2400 MOL, Belgium
2 Department of Earth system Sciences, University of California Irvine, CA 92697, United States
3 United Nations Food and Agriculture Organization, Viale delle Terme di Caracalla, 00153 Rome, Italy
4 Department Forest & Landscape, University of Copenhagen, Denmark
5 Institute for Environment and Sustainability, Joint Research Centre of the European Commission, I-21027 Ispra (VA), Italy

Abstract

Two pan-European forest maps were produced by the European Commission’s Joint Research Centre for the years 2000 and 2006. Both forest maps were derived from high-resolution, optical satellite imagery using an automatic processing technique, while the forest map from 2006 was further refined to map forest types using MODIS satellite imagery. This article provides a summary of the methodology and the associated accuracy assessment for the two maps.

Introduction

Forests are now considered to fulfill multiple-use roles including the supply of wood-based products, contribution to renewable energy resources, a role in landscape management and the mitigation of climate change. Consequently, there is an increasing need for timely, accurate information to the spatial distribution and composition of forest resources.

There have been several examples of large-scale satellite-based mapping projects that include a forest class. For instance, in Europe the Corine Land Cover project (EEA, 2006), on a global scale the Globcover/Globcorine dataset (Arino et al., 2008), pan-tropical monitoring of deforestation (Raši et al., 2011) and in the United States, the National LandCover Dataset (NLCD) (Stehman et al., 2008). While on a national scale, forest data have been derived from a combination of satellite imagery and national forest inventory data (Tomppo et al., 2008).

Despite these mapping efforts, there were no pan-European forest maps available that could be used to analyze the spatial composition and distribution of forests as well as determining the extent of forest fragmentation (Riitters et al., 2002). Such maps would also be required for carbon cycle modeling and inputs into ecological and forest fire danger models. As a result, the EC-JRC produced two forest maps with a spatial resolution of 25-meters for the reference years 2000 and 2006 from the classification of high spatial satellite imagery, such as image data with a spatial resolution between 20 and 30 meters.

Materials

About 415 Landsat-7 ETM+ image scenes were used to produce Forest Map 2000 (FMAP2000). The Forest Map 2006 (FMAP2006) and Forest Type Map 2006 (FTYP2006) were produced from IRS-LISS-3 scenes and additional SPOT4/5 scenes, mostly acquired in 2006. If conditions were sub-optimal due to clouds or unfavorable seasonality, images from 2005 and 2007 were also used. In addition, 16-day MODIS composites at a spatial resolution of 250 meters were used to produce the FTYP2006 dataset.

Ancillary data from the CORINE Land Cover (CLC) were used as training data for the classification process (Bossard et al., 2000). CLC includes 44 land cover and land use classes from which three correspond to forest classes including broadleaved, coniferous and mixed forests.

The accuracy assessment of the FMAP2000 was carried out using two datasets, the EU land use and land cover statistical survey (LUCAS) database and VISVAL (Pekkarinen et al., 2009), which was a reference dataset consisting of 5,193 validation points that were interpreted from Google Earth Imagery. The FMAP2006 was validated using the LUCAS datasets and the eForest Platform. The latter dataset represented a milestone in European forest data as it is a pan-European harmonized database consisting of data from over 1 million of national forest inventory (NFI) plots from 23 EU Member states. The information used for the accuracy assessment of the FMAP2006, included information on the forest type on a plot, broad age classification and percentage of crown cover recorded on the forest plot. Given that different NFI definitions and methodologies are used between countries, the eForest platform was harmonised to ensure consistency of NFI data between the participating countries (McRoberts et al. 2009).

Pre-processing and classification

The fine spatial resolution scenes (Landsat for FMAP2000, IRS LISS-3 and SPOT4/5 for FMAP2006 and FTYP2006) were geometrically corrected and projected to the European Terrestrial Reference System Lambert Azimuthal Equal Area projection (EPSG:3035). The data were also converted to top of atmosphere (TOA) radiances. Since no information on cloud cover was available, a cloud detection algorithm was implemented in house (Sedano et al. 2011). Cloud masks are of key importance for satellite derived products, in particular when produced in an operational processing chain. In case of the Moderate-resolution Imaging Spectroradiometer (MODIS), images were also corrected for atmosphere. A single cloud free composite was created for the near infrared band at 250 meter spatial resolution for each month. By stacking the composites for each month in 2006, a yearly product was obtained with 12 bands (Jan-Dec 2006).

The forest maps were classified with an automatic classification algorithm. The classifier was trained with the Corine Land Cover (CLC) map, which is a coarse land cover map of Europe with 44 land cover classes. It includes also forest classes, but due to its minimum mapping unit of 25 hectares, the small forest patches are not mapped. The lack of small forest patches is an issue, especially in Europe, where fragmented forests are frequent. The objective of producing FMAP2000 and FMAP2006 was to provide an alternative forest map at the best possible spatial detail.

The FMAP2000 was mapped using a k-Nearest Neighbor (k-NN) classifier (Finley and McRoberts, 2008). Two key improvements to the basic k-NN were suggested by Pekkarinen et al. (2009): Image segmentation prior to the classification step and an adaptive spectral representation analysis (ASRA) to improve the training process. The segmentation was used to speed up the k-NN classification, which is known to be inefficient for processing large data sets. The ASRA was introduced after clustering of segments into spectral classes. It seeks to identify representative combinations of spectral and informational classes using a contingency table, derived from the cluster labels and CLC classes. (For more details of the algorithm, the reader is referred to Pekkarinen et al., 2009).

The classification method for FMAP2006 and FTYP2006 was based on an artificial neural network (ANN) (Rumelhart, 1986). The ANN can provide accurate classification results (Chini, 2008) and once trained it is time efficient, therefore an image segmentation step was not needed.

The FTYP2006 introduces two forest types – broadleaved and coniferous forest. Due to the increased complexity of the classification problem multi-temporal information obtained from the MODIS sensor was added to the multispectral information. The temporal aspect of the spectral reflectance describes phenology, which is an indicator for land cover types (DeFries, 1994).

Map showing JRC Forest Map 2000. Source: EC-JRC – FRC Unit.

Figure 1: JRC Forest Map 2000. Source: EC-JRC – FRC Unit.

However, the MODIS data had a spatial resolution that is ten times coarser than the multispectral imagery. A data fusion method was therefore integrated in a two step classification approach (Kempeneers et al., 2011). In step one, the classifier created a forest map, classifying forest and non-forest only. In step two, a new classifier refined forest into forest types, excluding the non-forested pixels from the classification process. The multi-temporal data at medium spatial resolution were introduced only in step two. The idea is that, as the classes are refined, the complexity of the classification increases. At this point the classifier can benefit most from the added information from data fusion. The forest/non-forest map was mapped using only the spectral information at fine spatial resolution and therefore retained the finest spatial resolution possible.

Results

Map Product

FMAP2000 covers the EU-27 and Norway, Switzerland, Lichtenstein, Albania, Croatia, Macedonia, Montenegro, Serbia. The spatial extent of the FMAP2006 was the same, but also included Turkey. The maps included a forest, non-forest and water class. The FMAP2000, FMAP2006 and FTYP2006 are available for download at forest.jrc.ec.europa.eu/forestmap-download, and are presented in Figures 1 and 2, respectively.

Map showing JRC Forest Map 2006. Source: EC-JRC – FRC Unit.

Figure 2: JRC Forest Map 2006. Source: EC-JRC – FRC Unit.

Accuracy Assessment

The accuracy of the JRC forest maps was assessed using the three reference datasets outlined above. Namely, the VISVAL and LUCAS datasets were used for the FMAP2000, while the FMAP2006 was compared against the eFOREST and LUCAS datasets. The accuracy assessment statistics were produced for the entire maps (Table 1).

The overall accuracy (OA) of the FMAP2000 was 88.6 and 90.8 percent respectively for the VISVAL and LUCAS datasets, while the FMAP2006 was 88.0 and 84.0 percent for the eForest and LUCAS datasets (Table 1). The calculation of the Producer’s Accuracy (PA) related to the Error of Omission and User’s Accuracy (UA) related to the Error of Commission provided more insight into the performance of the forest maps.

Due to a combination of different input data, choice of classifiers and a distinct approach how classifiers were trained, the FMAP2000 and FMAP2006 show a different behavior. Especially in areas where forest is difficult to map such as dry vegetated areas in the Mediterranean region or peat bog in Scandinavia, results can differ substantially. In general, FMAP2006 was found to be more accurate in the forest that was mapped, at the price of underestimating the forest. This resulted in a higher PA of Forest for the FMAP2000 than for the FMAP2006. For instance, based on LUCAS, the PA was 83.9 percent for FMAP2000, compared to 66 percent for FMAP2006. On the other hand, the UA of Forest was found to be higher for FMAP2006 (87 percent based on eForest) than for FMAP2000 (78 percent based on VISVAL). It can be summarized that while the OA’s for both forest maps are comparable, the underestimation of the forest area was more pronounced in the FMAP2006 compared to FMAP2000.

Table showing FMAP2000 and FMAP2006 accuracies with respect to validation datasetsUsing the eForest dataset it was possible to produce an accuracy assessment at NUTS3 level (EC 2003) and to identify the spatial distribution of the class accuracies (e.g., PA of the class Forest). FMAP2006 performed best in central Europe (e.g. Germany, Austria and France), while the lowest accuracies occurred in Spain, Ireland and parts of Finland. The accuracy assessment of the FTYP2006 map is currently on-going with results expected in late 2012.

Conclusion

The approaches described in this paper demonstrate an operational system for large-scale satellite derived forest maps produced using automatic methods. The use of eForest platform as a reference dataset for FMAP2006 is a new and unique development in the assessment of large-scale, continental forest maps. The eForest platform represents the first attempt to harmonize the data from Member States National Forest Inventories. While the exercise demonstrated there was higher underestimation of the Forest class in FMAP2006, it was also established that there were some specific issues with the validation data. For example, the difficulty of full harmonization of NFI data due to differences in sampling designs and definitions of forest.

Future outlook

This year, the GMES Initial Operations (GIO) plans to produce high resolution land-cover layers including a forest map with a spatial resolution of 20 meters and 25 meters. This map will add a new data layer to the European forest map time-series.

References

Arino, O., Bicheron, P., Achard, F., Latham, J., Witt, R., & Weber, J.-L. (2008). Globcover: The most detailed portrait of Earth. European Space Agency Bulletin, 24–31.

Chini, M., Pacifici, F., Emer, W. J., Pierdicca, N., Del Frate, F. (2008) Comparing statistical and neural network methods applied to very high resolution satellite images showing changes in man made structures at rocky flats. IEEE Transactions on Geoscience and Remote Sensing. 46(6):1812

Bossard, M., Feranec, J. and Otahel, J. (2000) CORINE Land Cover Technical Guide — Addendum 2000, Technical report nr., 40 European Environment Agency.

DeFries, R. S., & Townshend, J. (1994) NDVI-derived land cover classifications at a global scale. International Journal of Remote Sensing 15(17):3567-3586.

EEA. The Thematic Accuracy of Corine Land Cover, Assessment Using LUCAS (Land Use/Cover Area Frame Statistical Survey); Technical Report No 7/2006 (2006) Available online: http://www.eea.europa.eu/publications/technical_report_2006_7/ (accessed on 12th February 2012).

Finley, A. & McRoberts, R. (2008) Efficient k-nearest neighbor searches for multi-source forest attribute mapping. Remote Sensing of Environment, 112(5):2203-2211. Available at: http://linkinghub.elsevier.com/retrieve/pii/S0034425708000175.

Kempeneers, P., Sedano, F., Seebach, L., Strobl, P. & San-Miguel-Ayanz, J. (2011) Data fusion of different spatial resolution remote sensing images applied to forest-type mapping. IEEE Transactions on Geoscience and Remote Sensing 49:4977-4986

EC (2003) Regulation (EC) No. 1059/2003 of the European Parliament and of the Council on the establishment of a common classification of territorial units for statistics (NUTS). Official Journal of the European Union.

McRoberts, R.E., Tomppo, E., Schadauer, K., Vidal, C., Stahl, G., Chirici, G., Lanz, A., Cienciala, E., Winter, S., & Smith, W. (2009) Harmonizing national forest inventories. Journal of Forestry, 107(4):179-187. Available at: https://www.scopus.com/inward/record.url?eid=2-s2.0-69849105679&partnerID=40&md5=2fc8544b4878ea4b70a271df5b96d7de.

Pekkarinen, A., Reithmaier, L. & Strobl, P. (2009) Pan-European forest/non-forest mapping with Landsat ETM+ and CORINE Land Cover 2000 data. ISPRS Journal of Photogrammetry and Remote Sensing, 64(2):171-183. Available at: http://linkinghub.elsevier.com/retrieve/pii/S0924271608001032.

Raši, R., Bodart, C., Stibig, H.-J., Eva, H., Beuchle, R., Carboni, S., Simonetti, D., Achard, F. (2011) An automated approach for segmenting and classifying a large sample of multi-date Landsat imagery for pan-tropical forest monitoring, Remote Sensing of Environment 115 (12):3659-3669. Available at: http://linkinghub.elsevier.com/retrieve/pii/S0034425711003312.

Riitters, K.H. Wickham, James D.; O’Neill, Robert V.; Jones, K. Bruce; Smith, Elizabeth R.; Coulston, John W.; Wade, Timothy G.; Smith & Jonathan H. (2002) Fragmentation of Continental United States Forests. Ecosystems, 5(8):815-822. Available at: http://www.springerlink.com/index/GQ9PY88BXGGUD4UY.pdf.

Rumelhart, D., & McClelland, J. (1986). Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 1: Foundations. MIT Press Cambridge, MA, USA.

Sedano, F., Kempeneers, P., Strobl, P., Kucera, J., Vogt, P., Seebach, L., San-Miguel-Ayanz, J. (2011) A cloud mask methodology for high resolution remote sensing data combining information from high and medium resolution optical sensors. ISPRS Journal of Photogrammetry and Remote Sensing 66(5):588–596


Stehman, S. V., Wickham, J. D., Wade, T. G., & Smith, J. H. (2008). Designing a multi-objective, multi-support accuracy assessment of the 2001 National Land Cover Data (NLCD 2001) of the conterminous United States, Photogrammetric Engineering & Remote Sensing 74: 1561-1571. Available at: http://cat.inist.fr/?aModele=afficheN&cpsidt=20971747.

Tomppo, E., Olsson, H., Ståhl, G., Nilsson, M., Hagner, O. & Katila, M. (2008) Combining national forest inventory field plots and remote sensing data for forest databases. Remote Sensing of Environment, 112(5):1982-1999. Available at: http://linkinghub.elsevier.com/retrieve/pii/S0034425708000242.

Pieter Kempeneers is an engineer in electronics and obtained his Ph.D. in physics on information extraction from hyperspectral imagery, applied to vegetation, from the University of Antwerp. Since 1999, he has been working as a researcher in remote sensing with the Flemish Institute for Technological Research (VITO), Belgium. From 2008 to 2011, he was with the Joint Research Centre of the European Commission at Ispra.