Developing a method to process satellite images for mapping coconut agrosystems
ISBN 978-85-88783-11-9
Authors
1Prune Christobelle, K.M.; 2Geo, C.D.; 3Frederic, B.; 4Sébastien, G.; 5Gaëlle, V.
1PHD, AIX-MARSEILLE UNIVERSITY, UMR ESPACE 7300 Email: prune-komaye@gmx.fr
2UMR 5175 CEFE / UMR AGAP Email: geo.coppens@cirad.fr
3CNRS-CIRAD UMR AMAP Email: borne@cirad.fr
4AIX-MARSEILLE UNIVERSITÉ, CNRS ESPACE UMR 7300 Email: sebastien.gadal@univ-amu.fr
5CNRS-CIRAD UMR AMAP Email: gaelle.viennois@ciard.fr
Abstract
Our study aims at developing generalizable method with the high resolution satellite images (VHR) for (i) mapping coconut-based agro-ecosystems, differentiating them from oil palm agrosystems and (ii) to count coconut trees. Two different environment are tested for the development of methodology: Vanuatu (heterogeneous landscape, very ancient plantations) and Ivory Coast (Marc Delorme Research Station, monoculture, regular spacing, oil palm plantations). We compared two methods of land use classification. The first one is similar to that described by Teina (2009), based on spectral analysis and Watershed segmentation, which we simplified by using the NDVI vegetation index. The second one is the semi-automatic classification based on texture analysis (PAPRI method of Borne, 1994). Their results were validated from manually digitized photo-interpretation maps. In both situations, the PAPRI method produces better results than that of Teina (global kappa of 0.60 vs. 0.40). Spectral signatures do not allow a sufficiently accurate mapping of coconut and do not differentiate it from oil palm, despite their different NDVI signatures. The PAPRI method differentiates productive coconut from mixed plantations and other vegetation, either high or low (70% accuracy). In both situations, Teina’s method allows counting 65% of the coconut trees when they are well spaced. To increase the method accuracy, we suggest (1) field surveys (for small scale studies) and/or finer image resolution, allowing a high precision in manual mapping with a better discrimination between coconut and oil palm, thus limiting the proportion of mixed pixels. (2) A phenological monitoring would be useful to distinguish coconut and oil palm agrosystems. (3) Hyper-spectral images allow extracting more precisely the respective signatures of both species. Another possibility would be (4) an object-oriented analysis as proposed by the eCognition software. Finally, (5) coupling the Lidar system with Watershed analysis would allow a better characterization of coconut varietal types.