Search

The present study presents an approach simultaneously using the spatial information and polarimetric information provided by RADARSAT-2 Quad Pol multipolarization image data for the classification of forest species. Two statistical models were used for classification purposes: (i) a Markov model taking into account the spatial statistical dependencies between adjacent sites based on an initial segmentation derived from the K-means algorithm and, (ii) a K distribution model using as parameters the covariance matrix containing all the polarimetric information and the shape parameter characteristic of the K distribution estimated using the moments of the theoretical and practical normalized intensities. The classification is optimized using the stochastic simulated annealing (SA) algorithm. Validation of the results was carried out through comparison with ground data observations. The variation of the backscattering coefficient σ° obtained for the RADARSAT-2 Quad Pol multipolarization images with incidence angles of 26° and 45° is equal to 3 dB for the different types of tree species stands. Using HH, VV, and HV linear polarizations it was possible to discriminate only four classes (watercourses, tolerant hardwoods, intolerant hardwoods, and conifers), with only a slight interclass difference of 1 dB. With a modification of the incidence angle from 26° to 45°, no significant change in the variation of the backscattering coefficient was noted in relation to the different types of tree species. The mean and overall precision results obtained for the classification are 81.47% and 79.12%, respectively, for the image with a 45° incidence angle and 77.13% and 72.35% for the image with a 26° incidence angle.[Traduit par la Rédaction]