www.postersession.com Fig 6 compares the RPAS aerial mosaic to the Landsat mosaic and derived tree cover maps [4] available from [3]. The GFW products have been converted from their lat/long projection into UTM zone 33S in the 20cm grid of the aerial mosaic. This might have introduced the significant shift in geolocation. Nevertheless, taking this colocation error into account, it seems that the 50% or 60% tree cover contour lines correspond best to the aerial field observations. The 40% tree cover line clearly overestimates the forest area in this case. Inside the EU FP7 ReCover project [1,2], optical and synthetic aperture radar (SAR) satellite- based forest maps, from RapidEye and ALOS Palsar data for the years 2011-2013 and 2010 respectively, were produced over a study site in Bandundu Province of the Democratic Republic of Congo (DRC) in support of REDD. A 10-day fieldtrip was carried out in the Mai- Ndombe district in March 2013 to collect forest inventory plots, digital photography and position of land cover types and transitions along a transect as well as airborne photos from a small remotely piloted aerial system (RPAS). The RPAS was a low-cost remote controlled quadrocopter “DJI Phantom” with an attached Canon Powershot S100 digital compact camera. Forest inventory plots are providing a very fine scale description of the vegetation giving important forest parameters, like tree species, stem diameter and estimated forest biomass, but are generally very labor intensive and time consuming resulting only in a very limited amount of useful data for validating medium and high resolution satellite-based land cover and forest maps. Transects with collected images and positions of land cover types and transitions, as well as aerial photography provide a larger data base for this purpose. The positional accuracy of forest/non-forest transitions from the ReCover products (Fig 1) as well as other freely available forest products from Global Forest Watch (GFW) [3,4] are assessed with this field data, considering also the vegetation density. A general forest and land cover description is also presented for this area in Fig 2. Methods Conclusions Validation of Optical and SAR Satellite-Based Forest Maps with Field Data in the Mai-Ndombe District, Dem. Rep. of Congo J.Haarpaintner 1 , M.Kohling 1 , F. Enssle 2 , P. Datta 2 , A. Mazinga 3 , J. Likunda 4 (1) Norut, P.O. Box 6434, Tromsø Science Park, N-9294 Tromsø (Norway), Email:joergh@norut.no (3) OSFAC, 14 Sergent Moke - Q/ Socimat, Concession Safricas, Ngaliema / Kinshasa (DRC) (2) Albert-Ludwigs University Freiburg, Tennenbacher Straße 4, D-79106 Freiburg (Germany) (4) MECNT/DIAF, Kinshasa (DRC) Bibliography 1. Häme, T., L. Sirro, E. Cabrera, J. Haarpaintner, J. Heinzel, J. Hämäläinen, B. de Jong, F. Paz Pellat, D. Pedrazzani, J. Reiche, “Chapter 8: ReCover: Services for the Monitoring of Tropical Forest to Support REDD+ “, in “Let’s Embrace Space, vol. II” by Schulte-Braucks, R., P. Breger, H. Bischoff, S. Borowiecka, S. Sadiq (eds.), Publication office of the European Union, ISBN 978- 92-79-22207-8, doi: 10.2769/31208, pp. 106-114, 2012. 2. Haarpaintner, J., M. Kohling, D. Pedrazzani, F. Enßle, and L. Mane, “Improving the Tropical Forest Remote Sensing Services for the Dem. Rep. of Congo inside the EU FP7 ''RECOVER'' Project”, Oral presentation at the ESA Living Planet Symposium 2013, Edinburgh, UK, 09-13 Sep. 2013. 3. http:/www.globalforestwatch.org 4. Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice, and J. R. G. Townshend. 2013. “High-Resolution Global Maps of 21st-Century Forest Cover Change.” Science 342 (15 November): 850–53. Data available on-line from: http://earthenginepartners.appspot.com/science-2013-global-forest. Abstract Results Study area and ReCover products Fig 1. Location and ReCover service area of ~68.000 km2 in DRC (a and b), the transect and location of field sites (with example from A and B)visited in March 2013 (c), and the satellite mosaics and derived forest/non- forest maps for RapidEye data from 2011-2013 (d and e) and ALOS Palsar FBD (RGB=HH,HV,NDI) from 2010 (f and g). Fig 2. Identified land cover classes during the field campaign: Primary tropical forest (a), Secondary tropical forest (b), Peat swamp forest (c), Inundated forest (d), Wet meadow (e), Savannah (f), Subsistence agriculture (g), Marshes (h). Aerial images were collected using a remote controlled quadrocopter DJI Phantom v.1 with an attached Canon Powershot S100 with GPS, logging the position coordinates of each photo. To be able to take continuously photos about every 3s, the Canon firmware has been hacked. The single aerial images were taken in shutter priority (TV) mode with a shutter speed of 1/1000s. They were post-processed into a geo-referenced image mosaics at a resolution of about 10 cm using Agisoft PhotoScan Professional. Point-clouds and digital surface models from multi-angle views could also be extracted. The geo-referenced mosaics are compared to satellite data, specifically high-resolution (5m) RapidEye scenes and 30m L-band ALOS Palsar. The position accuracy is evaluated using ground GPS logs from the transect to be about 5 m. Forest borders were manually drawn for comparison with FNF maps. Tree types can be easily identified from the aerial photos. 3D forest structure can be extracted from the point clouds and vegetation height can be compared to forest plot measurements. ACKNOWLEDGEMENTS Thanks to T. Häme (VTT) for leading the ReCover consortium. Satellite data has been provided through a GSC-DA grant from ESA, JAXA and RapidEye/BlackBridge. ReCover was funded by the European Commission EU FP7-SPACE-2010-1 program under grant agreement nr. 263075. M.Kohling’s position at Norut was funded by grant nr. 204430/E40 from the Norwegian Research Council. Thanks to the field team and the local logistical support by CDD A. Mazinga. a b c d e f g a b c a b Fig 3. DJI Phantom v1 with Canon Powershot S100 (a), DSM, height profile, and aerial mosaic (b) and 3D representation (c) over site A. Fig 5. 30m-resolution forest/non-forest border from the RapidEye (blue line) and ALOS Palsar (red line) classification overlaid over (a) the aerial mosaic (23.03.2013), (b) the RapidEye scene (1.8.2011) and (c) ALOS Palsar 2010 mosaic (RGB=HH,HV,NDI). a b c Fig 5 shows how the 30m- resolution forest/non-forest boarders from the RapidEye and ALOS Palsar classification correspond to the different data sets. There seems to be a one pixel shift between the RapidEye and ALOS FNF classification and RapidEye estimating more forest than ALOS Palsar. GPS tracks from the road taken during the transect shows that this shift results from a geolocation error in the RapidEye data. The full use of the 5m RapidEye resolution and a geolocation correction would probably show a much better performance for the RapidEye FNF classification. Measuring the average absolute distance between the FNF border from the satellite classification results and the one extracted manually from aerial photos of RPAS flights over enough study sites could potentially provide a better validation method of FNF maps than ordinary confusion matrixes from ground-based point sampling which do not take properly into account the spatial forest distribution. a b c Fig 6. The aerial mosaic 2013 (a) compared to the GFW Landsat mosaic from 2012 (b) and the GFW tree cover product from the year 2000 (c) Source: Hansen/UMD/Google/USGS/NASA [3,4]. The colored lines to road track (white), the manually extracted forest border (green) and 40% (yellow), 50% (red) and 60% (orange) tree cover contour lines. • Low cost RPAS systems can provide valuable field data that can probably better validate satellite remote sensing forest products/maps than work-intensive, single-point forest inventory plots at much lower operational costs. • RPAS observation can even provide forest structure and vegetation height over several hectares. • RapidEye data could potentially provide much better forest maps in full 5m-resolution, but in remote areas, its geo-location can be off in the order of 20 meters. • ALOS Palsar FNF are a good and weather independent alternative to optical data at 30m resolution. • The GFW tree cover product has been assessed showing that forest areas correspond to GFW tree cover above 50%. Fig 4. Single aerial image (a) and point cloud (b) from multi- angle observation over site B. A B