Contribution of Texture from TerraSar-X Radar Image for Forest Classification
Benelcadi, H1; Frison, P. L1; Lardeux, C2; Rudant, J.P1
1University Paris EST- Marne la vallée, FRANCE; 2ONF International, FRANCE
This study aims at evaluating the texture analysis of high spatial resolution images TerraSAR-X for mapping tropical forests. Tropical forests play an important role in improving the overall carbon footprint. Forests are a major carbon sink in the world. They store some 289 (Gt) of carbon in trees and vegetation. The amount of carbon stored in forest biomass, deadwood, litter and soil is greater than all the carbon in the atmosphere. Globally, carbon stocks in forest biomass decreased by about 0.5 Gt per year during the period 2005-2010, mainly due to a reduction of the total forest area (Global Forest Resources Assessment 2010). At the Conference of Parties to the UN Framework Convention on Climate Change (UNFCCC) in Montreal in 2005, tropical countries have launched an initiative to establish a compensation mechanism for Reducing Emissions from Deforestation and Degradation (REDD +), which could be fully implemented within the framework of a treaty on climate change post-2012. The implementation of REDD+ involves collecting detailed data of the aboveground forest biomass starting with the estimate in tones of dry matter per hectare and ending with the evaluation of socio-economic parameters of the project area. The launch of TerraSAR-X satellite operating in X band ( =3cm) with a metric spatial resolution in 2007, looks promising for the observation of tropical forests. Indeed, observations with smaller wavelength allows for finer spatial resolutions of about one meter. This feature provides access to textural information at scales of several meters, which was not accessible with the previously existing SAR sensors. This study evaluates the potential of the TerraSAR-X data, with a spatial resolution of 1 meter for the classification of tropical forests. Indeed, the contribution of the analysis of textural information for classification has been emphasized. The latter is understood through the analysis of Haralick textural parameters (Haralick, 1973). The retained algorithm of classification is the SVM (Support Vector Machine) (Lardeux, 2009), as it allows to take into account numerous parameters, which can be heterogeneous with respect to their physical dimension. Based on the SVM classification, an algorithm allowing to make a selection of the optimal textural parameters has been developed. The study site belongs to the forest of the Cardamom. It consists in a large forest area of 20 000 km ² in southern Cambodia; the region is subject to degradation and deforestation, as a consequence of the intense/ industrial farming of the land. Different types of vegetation were identified during a ground survey held in 2011 (primary forest, deciduous forest, bamboo, plantation, forest Melaluka, Savannah grass, and mangrove.). The data used for this study is a radar image acquired from the very high spatial resolution satellite TerraSAR-X, in spot light mode. It is an intensity single polarization VV image. The image size is 12,5 x 9,5 with a pixel size of 0.5 m and a spatial resolution of 1m. First results show that the addition of Haralick parameters to intensity channel may significantly improve the accuracy of the classification results from an overall accuracy of 37% to 80%. Using the method of feature selection “greedy forward”, we introduced the Haralick parameters computed in different range of distance from 1 to 80 pixels, the higher performance of classification are achieved with the higher distances of Haralick parameters. In addition, we noticed that small distances of Haralick parameters are penalized as it doesn’t give a higher performance of classification. Haralick parameters will be computed on another site (Cameroun) and a comparison will be made between different type of polarization (HH and VV) and spatial resolution (1m and 3m). Moreover other textural methods will be tested and analyzed. References: Global Forest Resources Assessment 2010, Main Report, United Nations of the Food and Agriculture Rome, 2010, 163, ISSN 1014-2894. Haralick R. M., Shanmugam K., Dinstein I., 1973.''Textural features for image classification''. IEEE Trans. Systems, Man Cybernetics, vol. smc-3, n°. 6, November. Lardeux C., Frison P.-L, Tison C., Souyris J .-C., Stoll B., Rudant J.-P., 2009 : Support Vector Machine for multi-frequency SAR polarimetric data classification. IEEE Trans. Geosc. Remote Sensing, vol. 47, n°12, 4143-4152. Lardeux C., Frison P.-L, Souyris J .-C., Tison C., Stoll B., Rudant J.-P., 2011 : Classification of tropical vegetation using multifrequency partial SAR polarimetry. IEEE Geosci. Rem. Sens. Letters, vol. 8, n° 1, pp. 133-137. Longépé N., Rakwatin P., Isoguchi O., Shimada M., 2008,”Assessment of AlOS PALSAR 50m FBD Data for Regional land cover classification by support vector machines''. Transactions on Geoscience and Remote Sensing. Van der Sanden J. J., 1997. ''RADAR Remote sensing to support tropical management''. Tropenbos-guyana series 5, Tropenbos-guyanaprogramme, Georgetown, Guyana. ISBN 90- 5484- 778- 1.