IEEE Proof IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008 1 Forest-Height Inversion Using High-Resolution P-band Pol-InSAR Data 1 2 Franck Garestier, Pascale C. Dubois-Fernandez, Senior Member, IEEE, and Isabelle Champion 3 Abstract—In this paper, a high-resolution P-band Pol-InSAR 4 data set acquired by the airborne RAMSES system over pine forest 5 stands of different height is investigated. A significant penetration 6 depth in all the polarimetric channels and a wide range of polari- 7 metric-phase-center heights are observed, attesting of an interac- 8 tion of the radar waves with different forest structural elements. 9 The main objective of this paper concerns forest-height inversion 10 at P-band. First, forest-modeling assumptions are evaluated us- 11 ing a priori information, such as ground-level and forest-height 12 measurements. The full extend of the forest height is shown to 13 be responsible of the volume decorrelation, and a significant ori- 14 entation effect is clearly identified over the highest stands. As a 15 consequence, the Oriented Volume over Ground model (OVoG) 16 is determined to be the most appropriated model for the 17 forest-height inversion. At P-band, the ground contribution is AQ1 18 present in all the polarimetric channels due to the important 19 penetration at this frequency. To overcome this difficulty, a 20 time–frequency optimization method based on sublook decom- 21 position is developed to separate the pure ground and canopy 22 contributions, allowing forest-height estimation with OVoG with 23 an rms error on the order of 2 m. In the last section of this paper, a 24 sensitivity analysis of the inversion with respect to two important 25 system parameters, the signal-to-noise ratio and the resolution, 26 is presented, leading to a discussion on the inversion robustness 27 in spaceborne conditions, where these system parameters are the 28 most deteriorated as compared to airborne configurations. 29 Index Terms—Forestry, interferometry, polarimetric radar, 30 radar imaging. 31 I. I NTRODUCTION 32 F OREST inventory and characterization are important for 33 environmental and climatic purposes. Radar imagery us- 34 ing spaceborne instruments presents great potential for for- 35 est investigation because of the possibility of global world 36 coverage with low weather sensitivity. P-band radar systems 37 allow us a deeper sampling of the vegetation layer than higher 38 frequency systems already present in space and constitute, 39 therefore, an interesting and promising tool for tree-height and 40 aerial-biomass assessment. Previous works have shown that the 41 backscattered signal received in the cross-polar channels can be 42 used for a direct estimation of the aerial biomass up to 120 t/ha 43 Manuscript received November 6, 2007; revised xxxx. This work was supported by CNES and ESA under the Egyptsar04 Contract. F. Garestier and P. Dubois-Fernandez are with the Research Unit, DEMR/ RIM, ONERA CSP, 13661 Salon Air Cedex, France (e-mail: fgarestier@ yahoo.com). I. Champion is with the Institut National de Recherche Agronomique (INRA)–Centre de Recherche Bordeaux, 33883 Villenave d’Ornon Cedex, France. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2008.922032 [1]–[5]. Advanced techniques, such as SAR polarimetry and 44 Pol-InSAR, have allowed the estimation of the forest height, 45 which can be linked to biomass through allometric equations 46 specific to a forest specie, at P-band over the Nezer pine forest 47 (Landes, France) [6], [7]. Height estimation using these two 48 last techniques appeared to be possible over the full range of 49 biomass found in the data set, from 0 to 150 t/ha, using a 50 supervised approach. In the first one of the proposed inversion 51 schemes, the polarimetric anisotropy is found to be linearly 52 correlated with the forest height, and the regression line derived 53 from in situ data provides the inversion relationship [8]. No 54 saturation of the anisotropy height is observed for this height 55 interval (0–25 m). The second technique is based on the Ran- 56 dom Volume over Ground model (RVoG) [9], [10], assuming 57 a known mean extinction coefficient (0.3 dB/m). The investi- AQ2 58 gation of inversion sensitivity to extinction has shown that an 59 error on this parameter does not impact drastically the inversion 60 performance, given a proper selection of the ambiguity height. 61 Both techniques have allowed a forest-height estimation with an 62 rms error not exceeding 2 m. In this paper, a fully unsupervised 63 Pol-InSAR inversion method is described, and the transposition 64 of the height inversion in spatial conditions is then investigated. 65 Test site and data are primarily presented in Section II. 66 The principal structural characteristics of the studied pine 67 forest are described, and references concerning dendrometric 68 measurements conducted on this forest along with allometric 69 relations linking its different structural parameters are cited. 70 The following part of the section concerns the description of 71 the system parameters associated with the P-band Pol-InSAR 72 data acquired by the RAMSES system [11], [12]. 73 Section III gives an overview of forest modeling adapted to 74 Pol-InSAR data used in this paper. RVoG [9], [10] assump- 75 tions are briefly recalled, and both effects of volume eleva- 76 tion [Elevated RVoG (ERVoG)] and particle orientation in the 77 volume layer [Oriented Volume over Ground model (OVoG)] 78 are addressed. In each case, inversion procedure and output AQ3 79 parameter robustness are discussed. 80 The next section concerns the Pol-InSAR data analysis. In- 81 terferometric coherence corresponding to elementary scattering 82 mechanisms is investigated over the different height forest 83 stands in order to characterize the associated scatterers and to 84 provide insights on P-band wave interaction with the forest. 85 Forest-height inversion is then considered in Section V. 86 Assumptions suitable for P-band forest modeling adapted to 87 Pol-InSAR analysis are first evaluated. The standard RVoG 88 technique relies on the assumption that one of the polarization 89 channels is linked to the pure volume contribution. Since this 90 is not the case at P-band, a time–frequency approach based 91 0196-2892/$25.00 © 2008 IEEE