This article was originally published in a journal published by Elsevier, and the attached copy is provided by Elsevier for the author’s benefit and for the benefit of the author’s institution, for non-commercial research and educational use including without limitation use in instruction at your institution, sending it to specific colleagues that you know, and providing a copy to your institution’s administrator. All other uses, reproduction and distribution, including without limitation commercial reprints, selling or licensing copies or access, or posting on open internet sites, your personal or institution’s website or repository, are prohibited. For exceptions, permission may be sought for such use through Elsevier’s permissions site at: http://www.elsevier.com/locate/permissionusematerial

Remote Sensing of Environment 109 (2007) 379 – 392 www.elsevier.com/locate/rse

py

Predicting and mapping mangrove biomass from canopy grain analysis using Fourier-based textural ordination of IKONOS images a

co

Christophe Proisy a,⁎, Pierre Couteron b , François Fromard c

Institut de Recherche pour le Développement (IRD), UMR AMAP, Boulevard de la Lironde, TA A51/PS2, Montpellier cedex 5, F-34398, France b IRD, UMR AMAP and Département d'Ecologie, Institut Français de Pondichéry (IFP), 11 St Louis Street, 605001 Pondicherry, India c Centre National de la Recherche Scientifique, UMR ECOLAB, 29 rue Jeanne Marvig, BP 4349, 31055 Toulouse cedex 4, France

al

Received 4 September 2006; received in revised form 16 January 2007; accepted 20 January 2007

on

Abstract

r's

pe

rs

Predicting structural organization and biomass of tropical forest from remote sensing observation constitutes a great challenge. We assessed the potential of Fourier-based textural ordination (FOTO) to estimate mangrove forest biomass from very high resolution (VHR) IKONOS images. The FOTO method computes texture indices of canopy grain by performing a standardized principal component analysis (PCA) on the Fourier spectra obtained for image windows of adequate size. For two distinct study sites in French Guiana, FOTO indices derived from a 1 m panchromatic channel were able to consistently capture the whole gradient of canopy grain observed from the youngest to decaying stages of mangrove development, without requiring any intersite image correction. In addition, a multiple linear regression based on the three main textural indices yielded accurate predictions of mangrove total aboveground biomass. Since FOTO indices did not saturate for high biomass values, predictions were furthermore unbiased, even for levels above 450 t of dry matter per hectare. Maps of canopy texture (with RGB coding) and biomass were then produced over 8000 ha of unexplored, low accessibility mangrove. Applying the FOTO method to the 4 m near-infrared channel yielded acceptable results with some limitations for characterization of juvenile mangrove types. We finally discuss the influence of technical aspects pertaining to VHR images and to FOTO implementation (especially the size of the window used to compute Fourier spectra) and we evoke the interesting prospect of broad regional validity offered by the method to characterize high biomass tropical forest from standardized measures of canopy grain. © 2007 Elsevier Inc. All rights reserved.

1. Introduction

th o

Keywords: Canopy texture; Mangrove; IKONOS; Very high resolution; Biomass; Tropical forest; French Guiana

Au

The structure and biomass of tropical forest must be known in order to describe the surface and energy exchange with the atmosphere and quantify carbon pools (e.g. Clark & Clark, 2000). Recent work argues for the use of large, multiple-forest datasets to improve the validity of biomass prediction from in-situ measurement (Chave et al., 2005). Within this context, the use of remote sensing has long been advocated for collecting data over large areas at low cost, although transferring predictive relations ⁎ Corresponding author. Tel.: +33 467 617 545; fax: +33 467 615 668. E-mail addresses: [email protected] (C. Proisy), [email protected] (P. Couteron), [email protected] (F. Fromard). 0034-4257/$ - see front matter © 2007 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2007.01.009

over space and time remains a major problem (Foody, 2003; Woodcock et al., 2001). Indeed, lack of in-situ atmospheric data hampers the calibration of optical images (e.g. Song & Woodcock, 2003), whereas both optical- and radar-derived intensity is unable to represent the whole range of structural aspect and biomass variation observable in tropical forest. Signal saturation is observed at the midlevel of leaf area index (Foody, 2003) and biomass range (e.g. Imhoff, 1995; Mougin et al., 1999), that commonly reach values of 8 m2 m− 2 (e.g. Ferment et al., 2001) and 500 t of dry matter per ha (e.g. Cummings et al., 2002) in tropical environments. Very high resolution (VHR) optical data such as the IKONOS or QUICKBIRD satellite images may deliver greater thematic accuracy (Mumby & Edwards, 2002) than that provided by SPOT or LANDSAT sensors (e.g. Morisette et al., 2003;

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

al

co

py

maps are urgently needed in many areas. Mangrove deserves specific effort made towards reliable and broad-scale estimation of its biomass (e.g. RAMSAR Convention on Wetlands, Kyoto and Carbon Initiative EORC/JAXA). In equatorial regions, mature mangrove has a biomass of up to 450 t of dry matter per ha (Fromard et al., 1998). Previous work aimed at predicting mangrove biomass from remotely sensed data (either optical or radar) have given unsatisfactory results for high biomass (Lucas et al., 2007; Mougin et al., 1999; Simard et al., 2006). The aim of this paper is to demonstrate that textural analysis applied to VHR IKONOS images can overcome previous limitations and provide reliable estimates and maps, even for high biomass mangrove. More precisely, we apply the Fourierbased Textural Ordination principle (Couteron et al., 2005, 2006) ‘FOTO’ method in this sequel to 1 m panchromatic and 4 m infrared IKONOS images acquired over two mangrove areas of French Guiana. Since this is the first application of the method to this type of data and to mangrove analysis, we assess the relative performance of panchromatic and infrared data and investigate the influence of the window size at the scale of which Fourier spectra are computed. Finally, we discuss the need, versatility, and potential of the FOTO method in the general study of forest biomass. 2. Study area and data collection

Au

th o

r's

pe

rs

Steininger, 2000). Their potential for predicting the structure of tropical forest has not yet been fully explored. Using metric pixels, tropical forest no longer appears as a continuous ‘green carpet’ (Asner et al., 2002), because the contrast between sunlit and shadowed canopy is fully explicit and potentially informative. However, visual interpretation of VHR satellite images (Asner et al., 2002; Clark et al., 2004) and intensity-based analysis (Thenkabail et al., 2004) were more or less efficient with regard to the study of forest structure and dynamics. Texture analysis can greatly help in maximizing VHR optical data predictions for forest attributes (e.g. Wulder et al., 1998). Extensive literature highlights the potential of a statistical approach to the characterization of image texture (e.g. Haralick, 1979) in forest canopies (e.g. Foody, 2003; Ingram et al., 2005; Kayitakire et al., 2006). Approaches based on Fourier analysis (e.g. Barbier et al., 2006), wavelet transforms (Ouma et al., 2006), or Gabor filters (Clausi & Deng, 2005; Jain & Farrokhnia, 1991) remain rather less investigated. In particular, their application to the study of tropical forests has not been sufficiently addressed. Textural analyses based on 2-D Fourier transforms have been useful in order to characterize and monitor vegetation or landscape features in tropical, semiarid areas (Barbier et al., 2006; Couteron, 2002; Couteron et al., 2006) but such studies did not attempt biomass prediction. Additionally, Couteron et al. (2005) used a Fourier-based index of canopy texture to predict, from a single aerial photograph, the variation of classical forest stand parameters such as tree diameter across more than 3500 ha of undisturbed, diversified tropical rainforest in French Guiana. Within the range of tree diameters for which an allometric equation is valid (Chave et al., 2004), such spatialized estimates could be also applied to a general allometric model of tree and stand biomass (Chave et al., 2005). There thus seems to be an intriguing potential for tropical forest biomass prediction through the textural analysis of VHR canopy images to produce large-scale datasets for rainforest features. For this purpose, there is one class of high biomass tropical forest for which reliable allometric models already exist; namely, the intertidal mangrove forest of intertropical regions. Mangrove is acknowledged for its diverse, worldwide environmental value (Saenger, 2003). The simpler floristic and structural characteristics of this environment in comparison with terra-firme rain forest have allowed the robust definition of relationships between tree diameter and aboveground biomass (e.g. Fromard et al., 1998; Sherman et al., 2003; Soares & Schaeffer-Novelli, 2005) for all the tree species comprising the growth stages. Additionally, from equations simulating tree diameter growth (Chen & Twilley, 1998), sophisticated models have been developed (e.g. Berger & Hildenbrandt, 2000, 2003) to account for biomass-density trajectories at local scale. Largescale contexts may also constrain the local pattern of structure and function of mangrove ecosystems (Ellison, 2002; Saenger & Snedaker, 1993). Since the ecosystem is fast changing due to both human activity (e.g. Valiela et al., 2001) and natural coastal change induced by hydro-oceanic forces (Blasco et al., 1996; Souza Filho et al., 2006), efficient, semiautomated methods to obtain frequent updates for biomass estimation and structure

on

380

2.1. French Guiana mangrove

Along the 350 km coastline of French Guiana, the spatial and temporal distribution of mangrove (Baltzer et al., 2004) is influenced primarily by sediment dispersal from the Amazon River. This mangrove region in particular has to face a succession of rapid and acute erosion and accretion phases resulting from the drift of giant mudbanks to the northwest (Allison & Lee, 2004; Gardel & Gratiot, 2005). The main growth stages are patchily distributed across the coastal landscape with clusters of young trees sometimes occurring in close vicinity to decaying areas. This very dynamic environment calls for the development of remote sensing-based methods to analyze broad-scale forest ecosystem dynamics, as demonstrated in Mougin et al. (1999), Proisy et al. (2000), and Fromard et al. (2004). As part of the National Program of Coastal Environment (PNEC) and the French Guiana Contract ‘Plan Etat-Région’, field experiments have been conducted since 2002 and two IKONOS images were obtained for two of our main study sites, as described below. 2.2. Site descriptions

The study sites, named Kaw (4° 45′ N, 52°5′ W) and Sinnamary (5°26′ N, 53°02′ W) measure 3 km × 12 km and 4 km × 11 km, respectively. Their topography is virtually flat. They are about 160 km apart, yet both were experiencing an accretion phase during the years of observation. Pioneer, young, mature and decaying growth stages were present. The overall differences in forest species composition are mainly observed in adult stands. The occurrence of Rhizophora trees in the lower

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

381

Table 1 Forest parameters collected in the two sites (KA: Kaw; SI: Sinnamary) Tree number

BrB

(cm)

(trees ha− 1)

(t DM ha− 1)

TrB

P M D P Y M A A Y P Y P MM D D Y D M Y D A Y M A D A

300 1200 800 300 200 1600 1600 1600 900 900 900 900 1200 10,000 10,000 1600 1800 600 200 800 300 675 675 1200 1800 1800

8.9 11.2 47.5 7.1 2.9 12.4 27.7 21.2 12.2 10.0 12.5 11.0 8.7 59.4 70.0 6.6 71.6 4.1 8.6 16.1 10.6 13.2 25.4 5.3 5.3 9.1

2867 1025 138 3533 15100 781 281 447 1700 2611 1967 2322 1367 63 53 2594 72 6400 1433 663 2000 2240 400 3141 5233 2300

18 48 58 38 42 45 66 35 19 23 28 7 34 40 54 26 59 39 11 42 23 22 34 35 46 24

Adjusted AGB

Timing difference (year)

106 255 346 123 130 271 241 254 127 125 158 113 241 296 436 105 464 252 62 282 150 123 199 220 319 139

al

72 195 276 78 75 204 170 210 102 95 122 26 200 252 377 66 399 206 47 234 120 96 159 179 266 109

AGB

96 249 341 106 125 254 215 213 83 73 100 91 201 267 409 47 Not required

py

Mean DBH

(m2)

co

Sampled area

on

KA2 KA4 KA5 KA8 KA9 KA10 KA11 KA12 KA13 KA14 KA15 KA16 KA17 KA18 KA19 KA20 SI1 SI2 SI3 SI4 SI5 SI6 SI9 SI11 SI12 SI13

Growth stage

rs

Plot ID.

1 1 1 1 1 1 3 3 3 3 3 3 3 4 4 4 0 0 0 0 0 0 0 0 0 0

pe

P, A, M, MM and D mean pioneer, adult, mature, mixed-mature and decaying growth stages. BrB, TrB and AGB denote branch, trunk and above-ground biomass, respectively. Adjusted AGB values refer to estimated values at the acquisition year of the IKONOS image over the Kaw region, i.e., 2001.

r's

stratum of Avicennia-dominated forest are infrequently observed in Sinnamary (Fromard et al., 2004) whereas they are regularly encountered in Kaw. At both sites, the typical species of flooded forest such as Symphonia globulifera, Virola surinamensis, Ficus sp. and the palm tree Euterpe oleracea are found in the lower stratum of mixed old Avicennia trees (Fromard et al., 1998). 2.3. Field dataset

Au

th o

Following visual analysis of both aerial photographs and IKONOS images of the study sites, forest areas for field sampling were selected with the aim of capturing the largest range of dynamic stages, regardless of the difficulties inherent in penetrating such inundated and dense forest. Field measurements were carried out from 2002 to 2005. Data were collected within areas ranging from 200 m2 for young stages with a high density of trees to 1 ha for mature and decaying lower density areas. The final forest dataset is composed of 16 and 10 plots for Kaw and Sinnamary, respectively (Table 1). Trunk diameter at breast height (DBH) and tree number per ha were assessed for trees of DBH N 1 cm belonging to the three mangrove species; i.e. Avicennia germinans, Rhizophora spp. (two species R. mangle and R. racemosa were considered together in the calculations), and Laguncularia racemosa. Total aboveground biomass (AGB) value and the associated trunk and branch fraction were computed from DBH measurements using a logarithmic allometric model between biomass (y) and DBH (x) of the form y = βxα (Fromard et al., 1998). It is noted

that these relationships were obtained from in-situ weighing biomass of trees with DBH b42, 32 and 9.6 cm for Avicennia, Rhizophora, and Laguncularia, respectively. The allometric relationship of tree biomass with larger DBH is thus less accurate. At the time of the field inventory, the AGB values ranged from 80 to 436 t of dry matter per ha (t DM ha− 1). 2.4. Biomass adjustment The problems of linking ground and remotely sensed data are numerous. The timing of ground data collection rarely corresponds with that for the image acquisition. In our case, forest ground truth observations were conducted from 2002 to 2005 (Table 1) and the IKONOS images were acquired in 2001 and 2003. Although most of the young stands were inventoried close to the IKONOS acquisition year, we judged it valuable to adjust ground-based estimates of biomass in the Kaw region. Here the timing difference with IKONOS acquisition was not more than 4 year. For this purpose, the FORMAN model (Chen & Twilley, 1998) was available to simulate the effects of soil characteristics on mangrove growth. We used the main modeling equation that assumes the optimal growth for a tree as given by:
 d½DBH GDBH:ð1−DBH:H=DBHmax :Hmax Þ ¼ dt 274 þ 3b2 :DBH−4b3 DBH2

ð1Þ

in which DBH is diameter at breast height (cm), H is the tree height (cm), DBHmax and Hmax are the maximum values for a

382

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

Table 2 IKONOS acquisition metadata Study sites

Acquisition date

Local time

Product type

Sun elevation

Sun azimuth

Incidence angle

Collection azimuth

Kaw Sinnamary

2001-08-26 2003-10-12

10:57 11:11

GeoTiff, 8 bits GeoTiff, 11 bits

66.2 69.2

74.4 126.8

7.2 16.3

58.3 86.4

Angles are expressed in degrees.

py

co

al

rs

2.5. IKONOS data

digitized aerial photographs (Couteron, 2002; Couteron et al., 2005, 2006). A given run of the method (Fig. 1) starts with the specification of a window size and the windowing of the image. The Fourier r-spectra are computed for the images and the variability between spectra is analyzed by principal components analysis (PCA). Window scores on the most prominent axes are then used as texture indices. The present work, which is the first application of the method to IKONOS images, has also developed a computing package (in Matlab® language), which allows the user to manage georegistered multisource data in ERMapper® native format, and to use interfaces to run analyses based on different study sites, sensors, and Fourier configurations (e.g. window sizes).

on

given tree species. G, b2 and b3 are species-specific constants as given in Table 1 of Chen and Twilley (1998). Other constants, DBHmax and Hmax, were locally calibrated from our own measurements in French Guiana. For Avicennia the limiting values were 125 cm and 42 m, for Rhizophora 60 cm and 38 m, and for Laguncularia 15 cm and 15 m. From adjusted estimates of tree DBH values obtained with reference to 2001, adjusted estimates of biomass (AGB) were computed for all plots of the Kaw region by re-applying the allometric relationship (Table 1). Maximum difference between field-derived and adjusted AGB values reached 58 t DM ha− 1 for plots KA15 and KA20 inventoried 3 and 4 year after image acquisition. However, since tree mortality cannot be estimated a posteriori with present knowledge of the ecosystem, it is important to note that the indirect figures for 2001 biomass are underestimated. Only the adjusted total trunk and branch biomass values were used in the following analyses.

Au

th o

r's

pe

The Kaw and Sinnamary regions were recorded by IKONOS in August 2001 and October 2003, respectively, in a period which corresponds to the dry season (Table 2). Our field observations confirmed that the soil was not flooded by fresh water during the dry season. The Sinnamary region was observed during low spring tide whereas the Kaw region was imaged 2 h after high tide. Although the effect of flooding may affect the texture analysis for sparse canopy marsh forests, it is considered of little influence in our case since all the mangrove areas we have studied displayed a closed canopy. The wavelength bands used were 0.45–0.93 μm (panchromatic), 0.45–0.52 μm (blue), 0.51–0.60 μm (green), 0.63–0.70 μm (red), and 0.76–0.85 μm (near infrared). Panchromatic images have a 1 m resolution whereas other channels are at 4 m. Both images were delivered in a georegistered UTM projection with a respective 8-bits and 11-bits radiometric depth GEOTIF format. Visual delineation of nonmangrove areas was performed in order to mask them before texture analysis (Fig. 1, step 1). This allowed us to exclude cloud, water, swamp, bare soils, and mud flat surfaces from textural analysis. In this paper, both panchromatic (PA) and near infrared (NIR) channels were analyzed. We selected infrared within the 4 m data spectral channels for its sensitivity to vegetation. The images were directly entered for analysis without any radiometric or geometric transformation. 3. The FOTO method The FOTO method is based on textural ordination, which has been initially designed for, and applied to the processing of

Fig. 1. Flow of operations to apply the FOTO method to biomass prediction from IKONOS data.

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

f = 1000.r.N−1.ΔS−1 where ΔS is the pixel size in meters. The discrete set of spatial frequencies can be also transformed into sampled wavelengths (in meters) as λ = 1000 / f. For 1 m panchromatic and 4 m near-infrared data, the Nyquist lowest resolvable frequencies are respectively 500 and 125 cycles per km.

3.1. 2-D fast Fourier transform and r-spectra For this study, window size (N) varying from 19 to 100 pixels, was used in the ‘windowing’ process (see also Discussion). Each square window from an initial image was subjected to the twodimensional discrete fast Fourier transform (FFT) (Fig. 1, step 2). This means that spectral radiance expressed in spatial domain by a function I(x,y) of image column (x) and row ( y), is transposed into the frequency domain as a function F( p,q), where p and q are spatial frequencies along the XY directions. The highest resolvable sampled frequency (Nyquist frequency) is N / 2, so 1 ≤ p ≤ N / 2 and 1 ≤q ≤N / 2. According to Mugglestone and Renshaw (1998), Fourier coefficients are defined from mean-corrected observations (Mxy = Ixy − bIN) as: N X N X x¼1

Mxy cos½2kðpx þ qyÞ=N 

ð1Þ

Mxy sin½2kðpx þ qyÞ=N :

ð2Þ

py

3.2. Textural ordination based on r-spectra

co

For each particular window size, r-spectra computed from the images of the Kaw and Sinnamary regions were grouped into a general table. Each row was the r-spectrum of a given window, whereas each column contained I(r) values; i.e. the portions of the variance of image radiance explained by a given spatial frequency or wavenumber, r. This table of spectra was further submitted to standardized PCA (Fig. 1, step 4), that is eigenvector analysis of the correlation matrix between spatial frequencies. The most prominent principal components were used as texture indices, which enabled us to ordinate windows along texture gradients.

y¼1

al

apq ¼ N −2

N X N X x¼1

y¼1

3.3. Texture mapping and biomass prediction

To display the results of canopy texture analysis and to illustrate its potential for mapping mangrove types, we coded PCA scores as composite red–green–blue images (red, green and blue colors express window scores against first, second and third axes, respectively). Such texture maps had a spatial resolution directly deriving from window size; i.e. in meters, Δs = N for panchromatic data and Δs = 4.N for near-infrared data. Hence, increasing the window size means decreasing the spatial resolution of output texture maps and erodes the study area on the fringe, since the windows partially composed of nonmangrove vegetation are not processed. This effect was particularly notable in Sinnamary where small areas of mangrove are surrounded by bare soil, whereas in Kaw, it was noted to occur along the Kaw River. To limit such an effect, we only considered window sizes smaller than 120 m; i.e. 100 and 30 pixels for panchromatic and near-infrared images, respectively.

Au

h

th o

r's

pe

rs

This can be interpreted as a convolution between image value and waveform of varying direction and spatial frequency. In practice, Fourier coefficients are computed in another way via the efficient two-dimensional FFT (Diggle, 1990). The subsequent step is to ignore phase information, for which the potential contribution to texture characterization remains a controversial question (Tang & Stewart, 2000) and to 2 2 compute periodogram values as Gpq = N− 2 (apq + bpq ). The periodogram can be equivalently expressed in polar form; i.e. with Irθ values, by referring to a waveform having spatial frequency r and traveling direction θ with r ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi (wavenumber) p2 þq2 and θ = tan− 1(p / q). The wavenumber r is defined as the number of times a signal repeats itself within the current window in direction θ (Mugglestone & Renshaw, 1998). Further simplifications can be applied to the periodogram by averaging on all possible traveling directions θ to yield an azimuthally averaged radial spectrum, a so-called ‘r-spectrum’ denoted by I(r): X lðrÞ ¼ ðkr2 Þ−1 ða2pq þ b2pq Þ; ð3Þ

on

and bpq ¼ N −2

383

where k is the number of periodogram values at spatial frequency r and σ2 is the image variance (Fig. 1, step 3). Images with a coarse texture tend to yield r-spectra that are skewed towards small spatial frequencies, while fine-textured images may be expected to determine spectra that are more balanced. Hence, although the r-spectrum neglects orientation information, it has proved useful in order to summarize textural properties related to coarseness/fineness, which are of great importance in dealing with canopies of natural forest (Couteron et al., 2005). Finally, to compare results from different window sizes we expressed spatial frequencies, f, in cycles per km, namely

Fig. 2. Percent of total variance explained by the first three principal component axes for window sizes ranging from 30 to 100 m and 76 to 200 m for 1 m panchromatic and 4 m near infrared images, respectively.

384

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

individual variability figures of less than 5%. For the infrared channel, ordination of r-spectra yielded similar results (Fig. 2). Contribution of the first axis to the total variance was higher for the 1 m panchromatic than for the 4 m near-infrared images (58% vs. 39%, on average for the 3 windows sizes). Conversely, the second axis was less explicative for PA images than for NIR images (13% vs. 25%, on average). The third axis contributed equivalently in both PA and NIR channels ordination and the contributions of subsequent axes were very low. Although the use of more than three principal components was unnecessary, further analyses are however required to determine the relative influence of both the spatial resolution of the image and canopy features on the main PCA axes (in terms of the meaning and accounted variance).

The final step (Fig. 1, step 5) was to relate ground truth forest biomass plots to FOTO indices using a linear model of the form: FP ¼ a0 þ

3 X

ac :Tc ;

ð4Þ

c¼1

co

py

where FP is a forest parameter (mainly biomass), and a0 and ac are the coefficients of the multiple regression of FP onto the Tc FOTO indices obtained from the first three PCA axes. The Tc indices were computed for each plot by averaging the FOTO indices values of all windows overlapping the area of dimension equal to the window size and centred on the plot. Inversion quality was characterized using the coefficient of determination r2, the root mean square error rmse expressed in arithmetic units and the relative error s expressed in percent.

4.2. Interpretation of FOTO ordination results for specific window sizes

4. Results

on

Au

th o

r's

pe

rs

For panchromatic (PA) data, the three first factorial axes of the PCA accounted for more than 78% of the total variability observed in r-spectra (Fig. 2), and this proportion was obtained using 30-pixel windows (i.e. 30 m) rising up to 85% using 100-pixel windows. For all window sizes in the range of 50 to 100 pixels, axes beyond PC3 proved to be of low interest with

We first discussed results of the FOTO algorithm obtained using 50 m windows (50 pixels) from 1 m PA data (Fig. 3). The first PCA axis expressed the opposition between fine and coarse texture corresponding to spatial frequencies of less than 100 cycles km− 1 (wavelengths λ N 10 m) and more than 250 cycles km− 1 (λ b 4 m), respectively. High loadings on the second axis were found for intermediate spatial frequencies and this determined a nonlinear dependence between axes 1 and 2 (and with axis 3, not shown). It is notable that plotting window

al

4.1. Main FOTO results with respect to data source and window size

Fig. 3. FOTO of IKONOS images, Kaw site: 1 m panchromatic image using 50-pixel windows (left graphs) and 4 m near infrared image using 30-pixel windows (right graphs). Correlation between PCA axes and spatial frequencies (in cycles per km) are shown in upper graphs whereas bottom figures present windows positions in the plane defined by axes 1 and 2.

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

385

from young to decaying forest plots, dominant frequencies decreased from 260 cycles km− 1 (λ = 3.8 m) to 35 cycles km− 1 (λ = 28 m). This suggests that r-spectra permit good discrimination of a wide array of forest structures. When using 120-m windows (30 pixels) from 4-m NIR data, the first PCA axis still expressed a difference between coarse and fine texture, which is between spatial frequencies under 40 cycles km− 1 and those above (Fig. 3). The second axis also corresponded to a complementary opposition between very coarse texture (less than 10 cycles km− 1) and intermediate textures in the range of 60 to 80 cycles km− 1. This second axis clearly separated in the windows displaying coarse canopy aspects, a group related to the decaying stage, and a further group

Au

th o

r's

pe

rs

on

al

co

py

scores against axes revealed a continuous coarseness–fineness gradient, expressing itself as a ‘horseshoe’ trend in the first PCA plane (Fig. 3, upper left graph). As the PCA axes are linearly independent by construction, they can safely be used in a multiple linear model (see below), notwithstanding the functional relationship of higher order observed. Mangrove development stages were each characterized by the dominance of a particular range of spatial frequencies as can be observed in averaged r-spectra for stands of equivalent growth stage (Fig. 4, r-spectra profiles). This texture gradient clearly agreed with the visual classification that could be made from canopy pictures (Fig. 4, square portions of images), whilst providing a quantitative and objective ordination of windows. Indeed,

Fig. 4. Averaged r-spectra for successive mangrove development stages. Particular examples of one hectare square windows over panchromatic (PA) and near infrared (NIR) images are displayed below. They correspond to stands KA13 (young), KA12 (adult), KA10 (mature) and SI1 (decaying).

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

Au

th o

r's

pe

rs

on

al

co

py

386

Fig. 5. Panchromatic-derived FOTO maps of canopy texture obtained by RGB coding on window scores on the three main PCA axes (red = PC1, green = PC2, blue = PC3). Map scales are identical for both Kaw and Sinnamary sites. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

387

Table 3 Coefficients of the regressions between the three texture indices and mangrove stand parameters (n = 26) Channel

N (pixels)

r2

rmse (t DM ha− 1)

s (%)

a0

a1

a2

a3

Total biomass

PA

30 50 75 100 19 25 30 30 50 75 100 19 25 30 30 50 75 100 19 25 30

0.91 0.92 0.89 0.87 0.87 0.80 0.70 0.91 0.92 0.88 0.86 0.85 0.79 0.69 0.62 0.62 0.61 0.60 0.74 0.62 0.59

34 33 39 42 43 53 64 31 30 35 38 40 48 57 10 10 10 11 9 10 11

16.9 17.3 20.5 23.7 24.5 24.0 31.0 18.4 19.3 23.7 27.5 29.4 28.1 36.3 24.4 25.3 25.4 25.3 21.9 24.6 26.7

183.2 188.1 185.6 187.1 196.7 204.4 215.2 144.8 149.2 146.9 148.4 155.3 162.2 172.2 32.9 33.3 33.1 33.2 36.0 36.7 37.5

− 25.8 − 19.6 − 15.4 − 12.8 − 44.1 − 32.9 − 31.9 − 23.1 − 17.5 − 13.6 −11.3 − 37.8 − 28.4 − 27.4 − 2.7 − 2.1 − 1.7 − 1.5 − 6.2 − 4.4 − 4.4

38.0 26.2 19.3 14.9 38.1 −10.3 −16.3 32.5 22.6 16.8 13.0 36.7 −10.7 −15.4 5.2 3.4 2.4 1.8 1.0 0.4 −0.1

− 25.6 − 19.7 − 13.7 − 7.9 11.8 36.5 14.9 − 26.1 − 19.6 − 13.6 − 7.9 13.8 36.4 16.9 − 0.2 − 0.6 − 0.5 − 0.3 − 1.5 0.0 − 1.9

PA

NIR

Branch biomass

PA

NIR

co

Trunk biomass

al

NIR

py

Forest parameter

canopy grain with scores along PC1 ranging from values 3 to 8 (Fig. 3). Overall, FOTO maps from near-infrared channels also captured adult growth stages consisting of broad tree crowns and intertree gaps (RGB maps not presented). However, the discrimination between young and adult mangrove areas was less precise than with FOTO maps derived from PA data.

4.3. Mapping FOTO textural indices

pe

rs

corresponding to mature mangrove. However, in the windows with fine-textured canopy (positive extremity of axis 1), the distinction between pioneer, young, and adult stages was no longer possible as shown in comparison of r-spectra (Fig. 4, top graphs) due to insufficient spatial resolution of the NIR images.

on

Coefficients of determination (r 2), root-mean square errors (rmse) in arithmetic units and relative error (s) in percent are indicated. Results with r 2 N 0.8 are presented in bold characters.

Au

th o

r's

From the above analyses, it was clear that at least two PCA axes should be considered to comprehensively depict canopy texture variation in the mangrove stands under study. To have simultaneous mapping we coded window scores against the three main PCA axes as RGB real values. Such FOTO maps, especially those deriving from panchromatic data (Fig. 5), proved greatly informative with a strong color contrast between texture types whatever the window size. For pioneer stages, high positive scores against the three axes (Fig. 3) yielded bright colors varying from black to yellow. Young stages were characterized by red–i.e. high scores on the only PC1 axis– whereas intergrades between blue and cyan corresponded to areas with adult trees (low positive scores on PC1 and negative scores on PC2). Green characterized decaying stages with high PC2 and very low PC1 scores. The RGB FOTO maps allowed us to capture, at a glance, the overall spatial organization of development stages within the sites and to get valuable insights on specific dynamics experienced by each site. In Kaw, there was a clear zonation of texture types perpendicular to the present location of the marine shoreline. The amount of color detail provided for each type of mangrove stage was also promising because it suggested both continuity and sensitivity of the texture indices. For example, areas featuring a high density of young trees showed a substantial variation in

4.4. Estimates of biomass from FOTO indices We estimated biomass from multiple linear regression models using the three textural indices (scores of the three main PCA axes) as independent variables. Results were compared according to the window size for both PA and NIR data (Table 3). The best results were obtained from the prediction of AGB values from PA data, with r2 above 0.87 and root mean square error (rmse) below 42 t DM ha− 1, whatever the window size. A similar quality of prediction was obtained for trunk biomass, although branch biomass appeared more difficult to predict (r2 of ca. 0.6). The most satisfactory results were gained using 50-pixel windows (50 m) for all three categories of biomass. Plotting predictions of AGB against field-measured values (Fig. 6) furthermore demonstrated that both of the study sites aligned with the same model with no obvious local bias around the line of slope 1. The relative errors did not exceed 24% and 28% of the field-estimated value for total and trunk biomass, respectively. In addition, estimates did not level off for high biomass values. To provide a synthetic overview of these results, we used the coefficients of the multiple-regression model to derive biomass maps from the three texture indices. Gray-scaled images were then produced by directly coding the range of biomass values (Fig. 7). The resulting maps agreed with our ground knowledge and

388

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

FOTO predictions from infrared images yielded less accurate, yet encouraging results for total and trunk biomass, with r2 N 0.79 for N = 19 or 25 pixels (Table 3). However, dispersion of points across the biomass range was slightly larger

Au

th o

r's

pe

rs

on

al

co

py

provided a reasonable interpolation of biomass information acquired from field plots. They undoubtedly confirmed the large potential of the FOTO method for deriving maps of aboveground biomass.

Fig. 6. Inversion relationships between ground-control biomass values and panchromatic (top graph) and infrared (bottom graph) FOTO-derived texture indices. Indices are computed from the PCA analysis of the pooled r-spectra tables of Kaw and Sinnamary sites.

389

pe

rs

on

al

co

py

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

Fig. 7. FOTO-derived biomass map of the Kaw site as estimated from the model given in Fig. 6.

r's

than it was using the panchromatic channel (Fig. 6) and the two sites yielded regression lines with seemingly distinct trends. 5. Discussion

th o

5.1. Pertinence of the FOTO method in mangrove studies

Au

The potential of the Fourier-based textural ordination (FOTO) of both 1 m and 4 m IKONOS images for analyzing and mapping mangrove canopy texture has been tested at two sites in French Guiana. Textural indices derived from the three main axes of the principal component analysis of Fourier r-spectra demonstrated their ability to capture the whole gradient of canopy grain observed from the youngest to the decaying stages of mangrove development. It is notable here that consistent results were drawn from two distinct images without prior radiometric correction, such as reflectance calibration, or histogram range concordance. With 1 m PA data, the entire texture gradient was captured whatever the scale of analysis defined by window size, whereas infrared 4 m data, although displaying a limitation with regard to the youngest stages, were able to properly capture the canopy grain variation from the adult to decaying stages. Following coding into RGB composite images, FOTO maps usefully revealed at a glance the spatial distribution of mangrove

types within a study site and showed contrast in terms of past and ongoing mangrove dynamics. Using the three textural indices as independent variables allowed us to reach accurate and unbiased predictions of both total aboveground and trunk biomass, as well as reasonable predictions for branch biomass. To our knowledge, this is the first time that independent variables derived from remote sensing have proved able to provide forest biomass predictions that did not level off for high levels. French Guiana mangrove P-band radar signatures displayed signal saturation for biomass approaching 250 t DM ha− 1 (Mougin et al., 1999). Fig. 6 shows no trend towards saturation and no obvious increase of variance estimation even for very high biomass values above 400 t DM ha− 1. FOTO maps appear able to provide information about the spatial distribution of biomass and mangrove types over thousands of unexplored hectares. FOTO maps appear able to provide information about the spatial distribution of biomass and mangrove types over thousands of unexplored hectares. However, we underline that allometric relationships predicting mangrove tree biomass from field measures (i.e. trees diameters) have to be improved and refined thanks to additional in-situ measurements including a wider range of diameters. This is a critical need to improve the accuracy of both biomass estimates in mature forest (Chave et al., 2005) and FOTO-derived AGB maps. We believe that such an objective is achievable in

390

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

understanding of the link between image aspect and forest structure, which is a critical issue for the more efficient use of remote sensing (e.g. Song et al., 2002). The robust statistical relationship found between canopy texture and forest parameters points towards a strong, yet poorly understood, coupling between canopy aspect and the threedimensional dynamical processes of vertical growth, intertree competition, and self-thinning. Further investigations on this fascinating topic could rely on a synergy between the FOTO method, which is able to characterize canopy aspect, and techniques providing information on the vertical distribution of biomass, such as Lidar (e.g. Gillespie et al., 2004) or radar (e.g. Quinones & Hoekman, 2004). The high-biomass equatorial mangrove regions present very favorable conditions for such investigations, since they grow in level coastal situations immune from relief effects and they are more cloud free than inland forest.

py

mangrove forests due to the little number of species. Besides contrasting stages of development with distinct canopy textures, e.g. young and decaying stands, may sometimes display similar biomass values (Proisy et al., 2002), which may also bias the FOTO estimation. Nevertheless, we believe that FOTO RGB texture maps can be very useful in the orientation of necessary field verification and will greatly contribute to research programs conducted on Amazonian mangrove forests.

al

5.4. Intersite calibrations and regional databases

on

Mangrove canopy grain is characterized by a texture gradient ranging from a wavelength of about 3 m for pioneer areas to 30 m for sparse old-growth stages. A comparable range of canopy texture (5 m to 50 m) was observed for diversified types of tropical terra firme forests in French Guiana (Couteron et al., 2005). Whereas very fine textures were found for extensive juvenile mangrove stands, fine textures were found for mature rainforests growing in steep areas of limited extent or for very small regeneration gaps that were not detectable even from VHR images. Consequently, the PCA yields distinct patterns, with a single axis attributable to canopy texture in the rainforest and mainly two axes in mangrove. Nevertheless, in both situations, FOTO-derived texture indices proved well correlated with fieldmeasured stand parameters, even though the considered forests were very different and notwithstanding distinct VHR initial material (digitized air photos vs. IKONOS images). Development of a standardized measurement for forest canopy grain with broad regional validity could be a realistic goal in the future. The first stage towards this objective would be to devise simple methods to calibrate r-spectra obtained from VHR images having various characteristics in acquisition parameters (viewing, solar position, etc.). The second step would be the establishment of regional databases featuring field-measured forest parameters and corresponding canopy images of adequate size for diversified and representative forest situations. From these basics, some kind of ‘meta’ FOTOanalysis could be run at regional scale in order to calibrate strong relationships between textural indices and forest parameters. Another reason to standardize texture indices would be to exploit the historical series of VHR data, namely digitized aerial photographs, in order to trace back the dynamics of forest ecosystems.

th o

r's

pe

rs

Windowing of images is the basic step in the FOTO method and the window size may have substantial influence on subsequent results. Small windows may be unable to adequately capture large canopy features observable in mature or decaying growth stages, whereas large windows may include features characterizing landforms or landscapes rather than canopy grain (Couteron et al., 2006). Using large windows also means decreasing the spatial resolution of the output FOTO maps, although the use of a sliding window, which is computationally intensive, can attenuate this effect. For this study, we considered ranges of window size that remained within reasonable boundaries, that is 30 m to 100 m for PA data and 76 m to 120 m for NIR data. (It is inadequate to work with windows smaller than 76 m with NIR data since a minimum number of pixels is needed for spectral analysis.) Although biomass prediction clearly indicated an optimal size of 50 m for PA and of 76 m for NIR, the results remain acceptable for all window sizes below 100 m. The principal component decomposition was, moreover, slightly improved with the use of 75-m or 100m windows. Further investigations are needed into the influence of VHR image technical characteristics on the results of the FOTO method. Indeed, radiometric and geometric corrections applied in IKONOS products conventionally available for sale could contribute to increased noise level (Ryan et al., 2003) and could bias Fourier r-spectra, especially for high frequencies relating to fine canopy grain. It would be appropriate to apply geometric corrections after FOTO processing of images rather than before. Accessibility to native ground sampled distance (GSD) format may thus be relevant.

co

5.2. Influence of window size and technical features of VHR data

Au

5.3. FOTO method for forest monitoring and modeling Both the present work and previous studies (Couteron et al., 2005, 2006) confirm the premise that swapping the spatial domain for the frequency domain is an efficient way to characterize vegetation and landscape spatial patterns using quantitative and objective coarseness/fineness gradients. In fact, although VHR images of mangrove and evergreen ‘terra firme’ rain forest do not suggest strict periodicity, the r-spectra of canopy images regularly exhibit ranges of dominant spatial frequencies, which can be used to characterize and discriminate forest types. Beyond this empirical finding, physical and biological determinants of dominant periodicity (or pseudoperiodicity) would be worth investigating to get a deeper

5.5. Conclusion Although Fourier transform and principal component analysis are common techniques, to our knowledge their combined use in monitoring vegetation is largely new. To compensate for the lack

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

r's

co

pe

rs

on

al

Acknowledgements The PNEC (Programme National en Environnement Côtier) and the French Guiana CPER (Contrat de Plan Etat-Région) supported acquisition of IKONOS data. We thank Antoine Gardel, Flora Lokonadinpoullé, Laurent Polidori, Didier Rochotte, Jean Louis Smock, Michel Tarcy and Adeline Thévand for aid in the field and constructive discussions. We are also grateful to Raphaël Pélissier and Jean-François Molino for their careful reading of the submitted manuscript. AMAP (Botany and Computational Plant Architecture) is a joint research unit which associates CIRAD (UMR51), CNRS (UMR5120), INRA (UMR931), IRD (2M123), and Montpellier 2 University (UM27), http://amap.cirad.fr. L'Institut Français de Pondichéry (IFP) is a research institute sponsored by the French Ministry of Foreign Affairs, http://www.ifpindia.org. ECOLAB (Laboratory of Functional Ecology) is a joint research unit which associates CNRS (UMR 5245) and Paul Sabatier University of Toulouse (UM27), http://www.ecolab.ups-tlse.fr.

Burrus, C. S., Gopinath, R. A., & Guo, H. (1997). Introduction to wavelets and wavelets transforms: A primer. New Jersey: Prentice Hall0-13-489600-9. Chave, J., Andalo, C., Brown, S., Cairns, M. A., Chambers, J. Q., Eamus, D., et al. (2005). Tree allometry and improved estimation of carbon stocks and balance in tropical forests. Oecologia, 145, 87−99. Chave, J., Condit, R., Aguilar, S., Hernandez, A., Lao, S., & Perez, R. (2004). Error propagation and scaling for tropical forest biomass estimates. Philosophical Transactions of the Royal Society of London. Series B, 359, 409−420. Chen, R., & Twilley, R. R. (1998). A gap dynamic model of mangrove forest development along gradients of soil salinity and nutrient resources. Journal of Ecology, 86, 37−51. Clark, D. B., Castro, C. S., Alvarado, L. D. A., & Read, J. M. (2004). Quantifying mortality of tropical rain forest trees using high-spatialresolution satellite data. Ecology Letters, 7, 52−59. Clark, D. B., & Clark, D. A. (2000). Landscape-scale variation in forest structure and biomass in a tropical rain forest. Forest Ecology and Management, 137, 185−198. Clausi, D. A., & Deng, H. (2005). Design-based texture feature fusion using Gabor filters and co-occurrence probabilities. IEEE Transactions on Image Processing, 14, 925−936. Couteron, P. (2002). Quantifying change in patterned semi-arid vegetation by Fourier analysis of digitized aerial photographs. International Journal of Remote Sensing, 23, 3407−3425. Couteron, P., Barbier, N., & Gautier, D. (2006). Textural ordination based on Fourier spectral decomposition: A method to analyze and compare landscape patterns. Landscape Ecology, 21, 555−567. Couteron, P., Pélissier, R., Nicolini, E., & Paget, D. (2005). Predicting tropical forest stand structure parameters from Fourier transform of very highresolution remotely sensed canopy figures. Journal of Applied Ecology, 42, 1121−1128. Cummings, D. L., Boone Kauffman, J., Perry, D. A., & Flint Hughes, R. (2002). Aboveground biomass and structure of rainforests in the southwestern Brazilian Amazon. Forest Ecology and Management, 163, 293−307. Diggle, P. J. (1990). Time series. A biostatistical introduction. New York: Oxford University Press0-19-852226-6. Ellison, A. M. (2002). Macroecology of mangroves: Large-scale patterns and processes in tropical coastal forests. Trees-Structure and Function, 16, 181−194. Ferment, A., Picard, N., Gourlet-Fleury, S., & Baraloto, C. (2001). A comparison of five indirect methods for characterizing the light environment in a tropical forest. Annals of Forest Science, 58, 877−891. Foody, G. M. (2003). Remote sensing of tropical forest environments: Towards the monitoring of environmental resources for sustainable development. International Journal of Remote Sensing, 24, 4035−4036. Fromard, F., Puig, H., Mougin, E., Marty, G., Betoulle, J. L., & Cadamuro, L. (1998). Structure, above-ground biomass and dynamics of mangrove ecosystems: New data from French Guiana. Oecologia, 115, 39−53. Fromard, F., Vega, C., & Proisy, C. (2004). Half a century of dynamic coastal change affecting mangrove shorelines of French Guiana. A case study based on remote sensing data analyses and field surveys. Marine Geology, 208, 265−280. Gardel, A., & Gratiot, N. (2005). A satellite image-based method for estimating rates of mud bank migration, French Guiana, South America. Journal of Coastal Research, 21, 720−728. Gillespie, T. W., Brock, J., & Wright, C. W. (2004). Prospects for quantifying structure, floristic composition and species richness of tropical forests. International Journal of Remote Sensing, 25, 707−715. Haralick, R. M. (1979). Statistical and structural approaches to texture. Proceedings of the IEEE, 67, 786−804. Imhoff, M. L. (1995). Radar backscatter and biomass saturation: Ramifications for global biomass inventory. IEEE Transactions on Geoscience and Remote Sensing, 33, 511−518. Ingram, J. C., Dawson, T. P., & Whittaker, R. J. (2005). Mapping tropical forest structure in southeastern Madagascar using remote sensing and artificial neural networks. Remote Sensing of Environment, 94, 491−507. Jain, A. K., & Farrokhnia, F. (1991). Unsupervised texture segmentation using Gabor filters. Pattern Recognition, 24, 1167−1186.

py

of established software, we created an easy-to-use interface under Matlab®, which allows the processing of any type of image, the choice of various FOTO configurations and the display of texture indices as RGB maps. Additional details are provided on http:// amapmed.free.fr/FOTO/. Requests for the use of the FOTO method via this computing interface will be welcome. Prospects for further development of this approach may be considered. The concept of textural ordination, on which FOTO relies, may also be based in the future on other multivariate methods for ordination and importantly, or other techniques of signal decomposition, such as wavelets or local trigonometric bases (Burrus et al., 1997). Textural features expressing local properties of the image (e.g. Ingram et al., 2005; Kayitakire et al., 2006) may also be used to complement Fourier spectra and be added to the table of spectra before ordination. This would be a way to explore further the potential of feature fusion techniques (Clausi & Deng, 2005) in the field of forest survey and vegetation monitoring.

th o

References

Au

Allison, M. A., & Lee, M. T. (2004). Sediment exchange between Amazon mudbanks and shore-fringing mangroves in French Guiana. Marine Geology, 208, 169−190. Asner, G. P., Palace, M., Keller, M., Pereira, R., Silva, J. N. M., & Zweede, J. C. (2002). Estimating canopy structure in an Amazon Forest from laser range finder and IKONOS satellite observations. Biotropica, 34, 483−492. Baltzer, F., Allison, M., & Fromard, F. (2004). Material exchange between the continental shelf and mangrove-fringed coasts with special reference to the Amazon–Guianas coast. Marine Geology, 208, 115−126. Barbier, N., Couteron, P., Lejoly, J., Deblauwe, V., & Lejeune, O. (2006). Selforganized vegetation patterning as a fingerprint of climate and human impact on semi-arid ecosystems. Journal of Ecology, 94, 537−547. Berger, U., & Hildenbrandt, H. (2000). A new approach to spatially explicit modelling of forest dynamics: Spacing, ageing and neighbourhood competition of mangrove trees. Ecological Modelling, 132, 287−302. Berger, U., & Hildenbrandt, H. (2003). The strength of competition among individual trees and the biomass-density trajectories of the cohort. Plant Ecology, 167, 89−96. Blasco, F., Saenger, P., & Janodet, E. (1996). Mangroves as indicators of coastal change. Catena, 27, 167−178.

391

C. Proisy et al. / Remote Sensing of Environment 109 (2007) 379–392

al

co

py

Sherman, R. E., Fahey, T. J., & Martinez, P. (2003). Spatial patterns of biomass and aboveground net primary productivity in a mangrove ecosystem in the Dominican Republic. Ecosystems, 6, 384−398. Simard, M., Zhang, K. Q., Rivera-Monroy, V. H., Ross, M. S., Ruiz, P. L., Castaneda-Moya, E., et al. (2006). Mapping height and biomass of mangrove forests in Everglades National Park with SRTM elevation data. Photogrammetric Engineering and Remote Sensing, 72, 299−311. Soares, M. L. G., & Schaeffer-Novelli, Y. (2005). Above-ground biomass of mangrove species. I. Analysis of models. Estuarine, Coastal and Shelf Science, 65, 1−18. Song, C., & Woodcock, C. E. (2003). A regional forest ecosystem carbon budget model: Impacts of forest age structure and landuse history. Ecological Modelling, 164, 33−47. Song, C., Woodcock, C. E., & Li, X. (2002). The spectral/temporal manifestation of forest succession in optical imagery: The potential of multitemporal imagery. Remote Sensing of Environment, 82, 285−302. Souza Filho, P. W. M., Farias Martins, E. D. S., & Da Costa, F. R. (2006). Using mangroves as a geological indicator of coastal changes in the Braganca macrotidal flat, Brazilian Amazon: A remote sensing data approach. Ocean & Coastal Management, 49, 462−475. Steininger, M. K. (2000). Satellite estimation of tropical secondary forest aboveground biomass: Data from Brazil and Bolivia. International Journal of Remote Sensing, 21, 1139−1157. Tang, X., & Stewart, W. K. (2000). Optical and sonar image classification: Wavelet Packet Transform vs Fourier Transform. Computer Vision and Image Understanding, 79, 25−46. Thenkabail, P. S., Enclona, E. A., Ashton, M. S., Legg, C., & De Dieu, M. J. (2004). Hyperion, IKONOS, ALI, and ETM+ sensors in the study of African rainforests. Remote Sensing of Environment, 90, 23−43. Valiela, I., Bowen, J. L., & York, J. K. (2001). Mangrove forests: One of the world's threatened major tropical environments. Bioscience, 51, 807−815. Woodcock, C. E., Macomber, S. A., Pax-Lenney, M., & Cohen, W. B. (2001). Monitoring large areas for forest change using Landsat: Generalization across space, time and Landsat sensors. Remote Sensing of Environment, 78, 194−203. Wulder, M. A., Ledrew, E. F., Franklin, S. E., & Lavigne, M. B. (1998). Aerial image texture information in the estimation of northern deciduous and mixed wood forest leaf Area index (LAI). Remote Sensing of Environment, 64, 64−76.

Au

th o

r's

pe

rs

Kayitakire, F., Hamel, C., & Defourny, P. (2006). Retrieving forest structure variables based on image texture analysis and IKONOS-2 imagery. Remote Sensing of Environment, 102, 390−401. Lucas, R. M., Mitchell, A. L., Rosenqvist, A., Proisy, C. A. M., & Ticehurst, C. (2007). The potential of L-band SAR for quantifying mangrove characteristics and change: Case studies from the Tropics. Aquatic Conservation: Marine and Freshwater Ecosystems. Special issue "Satellite-based radar-developing tools for wetlands management, 17(3). Morisette, J. T., Nickeson, J. E., Davis, P., Wang, Y., Tian, Y., Woodcock, C. E., et al. (2003). High spatial resolution satellite observations for validation of MODIS land products: IKONOS observations acquired under the NASA Scientific Data Purchase. Remote Sensing of Environment, 88, 100−110. Mougin, E., Proisy, C., Marty, G., Fromard, F., Puig, H., Betoulle, J. L., et al. (1999). Multifrequency and multipolarization radar backscattering from mangrove forests. IEEE Transactions on Geoscience and Remote Sensing, 37, 94−102. Mugglestone, M. A., & Renshaw, E. (1998). Detection of geological lineations on aerial photographs using two-dimensional spectral analysis. Computers & Geosciences, 24, 771−784. Mumby, P. J., & Edwards, A. J. (2002). Mapping marine environments with IKONOS imagery: Enhanced spatial resolution can deliver greater thematic accuracy. Remote Sensing of Environment, 82, 248−257. Ouma, Y. O., Ngigi, T. G., & Tateishi, R. (2006). On the optimization and selection of wavelet texture for feature extraction from high-resolution satellite imagery with application towards urban-tree delineation. International Journal of Remote Sensing, 27, 73−104. Proisy, C., Mougin, E., Fromard, F., & Karam, M. A. (2000). Interpretation of polarimetric radar signatures of mangrove forests. Remote Sensing of Environment, 71, 56−66. Proisy, C., Mougin, E., Fromard, F., Trichon, V., & Karam, M. A. (2002). On the influence of canopy structure on the polarimetric radar response from mangrove forest. International Journal of Remote Sensing, 23, 4197−4210. Quinones, M. J., & Hoekman, D. H. (2004). Exploration of factors limiting biomass estimation by polarimetric radar in tropical forests. IEEE Transactions on Geoscience and Remote Sensing, 42, 86−104. Ryan, R., Baldridge, B., Schowengerdt, R. A., Choi, T., Helder, D. L., & Blonski, S. (2003). IKONOS spatial resolution and image interpretability characterization. Remote Sensing of Environment, 88, 37−52. Saenger, P. (2003). Mangrove ecology, silviculture and conservation. Dordrecht, NL: Kluwer Academic Publishers1-4020-0686-1. Saenger, P., & Snedaker, S. C. (1993). Pantropical trends in mangrove aboveground biomass and annual litter fall. Oecologia, 96, 293−299.

on

392