Estimation of tree canopy cover in evergreen oak woodlands using ...

Forest Ecology and Management 223 (2006) 45–53 www.elsevier.com/locate/foreco

Estimation of tree canopy cover in evergreen oak woodlands using remote sensing Joaõ M.B. Carreiras *, Jose´ M.C. Pereira, Joaõ S. Pereira Department of Forestry, Instituto Superior de Agronomia, Tapada da Ajuda, 1349-017 Lisboa, Portugal Received 1 April 2005; received in revised form 21 October 2005; accepted 24 October 2005

Abstract The montado/dehesa landscapes of the Iberian Peninsula are savannah-type open woodlands dominated by evergreen oak species (Quercus suber L. and Q. ilex ssp. rotundifolia). Scattered trees stand over an undergrowth of shrubs or herbaceous plants. To partition leaf area index between trees and the herbaceous/shrubby understorey requires good estimates of tree canopy cover and is of key importance to understand the ecology and the changes in land cover. The two vegetation components differ in phenology as well as in radiation and rainfall interception, water and CO2 fluxes. The main goal of this study was to estimate tree canopy cover in a montado/dehesa region of southern Portugal (Alentejo) using remote sensed data. For this purpose we developed empirical models combining measurements obtained through the analysis of aerial photos and reflectance from Landsat Thematic Mapper (TM) individual channels, vegetation indices, and the components of the Kauth–Thomas (K–T) transformation. A set of 142 plots was designed, both in the aerial photos and in the satellite data. Several simple and multiple linear regression models were adjusted and validated. A subset of 75% of the data (n = 106) was used for model fitting, and the remainder (n = 36) was used for model assessment. The best linear equation includes Landsat TM channels 3, 4, 5 and 7 (r2 = 0.74), but the Normalised Difference Vegetation Index (NDVI), the components of the K–T transformation, and the Atmospherically Resistant Vegetation Index (ARVI) also performed well (r2 = 0.72, 0.70, and 0.69, respectively). The statistics of prediction residuals and tests of model validation indicates that these were also the models with better predictive capability. These results show that detection of low/medium tree canopy cover in this type of land cover (i.e. evergreen oak woodlands) can be accomplished with the help of high and medium spatial resolution satellite imagery. # 2005 Elsevier B.V. All rights reserved. Keywords: Landsat Thematic Mapper (TM); Aerial photo; Linear regression; Evergreen oak woodlands; Tree canopy cover

1. Introduction Savannah-type Mediterranean evergreen oak woodlands are widely distributed in the Iberian Peninsula, as well as in other areas with the Mediterranean type of climate, e.g., in parts of California, Chile, South Africa and Australia. These are complex ecosystems with open, heterogeneous canopies with shrub or annual herbaceous understories (Joffre et al., 1999; Baldocchi et al., 2004). In the Iberian Peninsula, large areas of these woodlands depend upon human action, forming a multiple use agroforestry system, called montado in Portugal and dehesa in Spain. For example, in Portugal, they cover an area of approximately 1.2 Mha (DGF, 2001). The most common oak species are holm oak (Quercus rotundifolia Lam. syn. Quercus ilex L. ssp. rotundifolia) and cork oak (Quercus suber L.). They

* Corresponding author. Tel.: +351 213653387; fax: +351 213645000. E-mail address: [email protected] (Joaõ M.B. Carreiras). 0378-1127/$ – see front matter # 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.foreco.2005.10.056

are exploited for cork, wood and extensive agriculture or grazing, in proportions that vary with local conditions and history. The landscapes are typically heterogeneous and include three distinct variants, which differ in terms of land use intensity and land cover structure (Blanco et al., 1997): - Relatively denser oak woodlands are more common in steep areas and poor soils, unsuitable for agricultural use. Land use intensity is low, and a diverse shrub community species dominates the understorey. In Portugal dense cork oak stands were planted for cork production in better soils. - Pastures represent an intermediate level of land use intensity. Tree cover is sparser, and the understorey is dominated by a great variety of annual/biennial herbaceous plant communities. The productivity of these communities is dominated by climatic factors, and typically they display large interannual variability in plant cover. - Dryland farming of cereal crops in areas with deeper soils, such as the bottom of valleys. During fallow years, these areas

46

J.M.B. Carreiras et al. / Forest Ecology and Management 223 (2006) 45–53

may be used as pastures. There is a tendency for abandonment of this type of agriculture and the montado/dehesa reverts to one of the former land cover types (Lourenço et al., 1998). Tree density in montado ecosystems frequently ranges from 30 to 60 trees ha1 (Blanco et al., 1997), but higher and lower densities are common. In Portugal, about half of the area of Q. suber montados has tree canopy cover in the 10–30% class, one quarter of the area in the 30–50% class, and the remaining quarter in the >50% tree canopy cover class, whereas holm oak woodlands are even sparser, with about 85% of the area in the 10–30% canopy cover class (DGF, 2001). These low tree densities result from specific land use objectives and management practices, which transformed the relatively denser primordial woodlands and forests into landscapes structurally similar to savannas. The Mediterranean evergreen oak woodlands differ from tropical savannahs in the timing of rainfall, which in the former occur during the relatively cool autumn–winter months. This leads to severe water deficits during the hot Mediterranean summer. As a consequence large intra-annual variations in plant cover occur (Baldocchi et al., 2004). The understanding of ecosystem structure and function (e.g., radiation and rainfall interception, water and CO2 fluxes) requires the partitioning of total leaf area index between the tree component and the herbaceous/shrubby understorey. The two vegetation layers differ in phenology and functioning. For example, whereas the sparse tree cover remains green and may have access to perennial water sources for year-long metabolic activity (David et al., 2004), the herbaceous layer dries out in the dry summer. In addition to the spatial and temporal heterogeneity, the montado/dehesa landscapes are today subject to rapid changes resulting from land abandonment, changing fire regimes and tree decline (Brasier, 1992, 1996; Terradas, 1999; Stoate et al., 2001; Romero-Calcerrada and Perry, 2004). Given the economical, ecological, and historical significance of these ecosystems, it is important to monitor the changes that affect them, in a timely and cost-effective manner. Satellite remote sensing provides a potentially useful tool for such monitoring, but the problem is complex because of the intrinsic diversity, structural complexity, and seasonal variability of the montado/dehesa landscapes. The designation of a relatively broad range of land cover types, under a common land use name, leads to a problem of mismatch between informational classes (i.e. those meaningful to a user), and spectral classes (those which share similar spectral properties) (Swain and Davies, 1978). In the case of the montado/dehesa systems, informational land use classes tend to be very heterogeneous from the spectral standpoint, reflecting their internal ecological variability. This is a severe problem for qualitative land use mapping, where the goal is to discriminate between distinct land cover types. Quantitative characterisation of a biophysical attribute such as tree canopy cover is also quite challenging in these landscapes, because of the very broad range in tree density, and of the wide variety of understorey cover types. The signal (tree canopy) has to be detected and

quantified against a very noisy background of bare soils, agricultural crops, pastures and shrubs. The main goal of this study was to estimate tree canopy cover in a montado/dehesa region of southern Portugal (Alentejo) using remote sensed data. Several studies have produced good and promising results. Joffre and Lacaze (1993) used panchromatic SPOT High Resolution Visible (HRV) (10 m spatial resolution) and Landsat Thematic Mapper (TM) (30 m spatial resolution) satellite imagery to estimate tree density in a savannah-type ecosystem (holm and cork oaks), in southern Spain; tree density was estimated with reasonable accuracy through the application of spatial filters to the panchromatic SPOT HRV band (r = 0.94), from the original panchromatic SPOT HRV (r = 0.85), and using a combination of Landsat TM channels 2 (TM2), 3 (TM3) and 4 (TM4) (r = 0.61). Salvador and Pons (1998) used Landsat TM imagery combined with fieldwork to estimate dendrometric variables (including tree canopy cover) used in forest inventories in the Mediterranean region, relying on regression models; these were found to be consistent with the expected vegetation spectral response; most of the multiple linear regression models fit well the data, allowing quantitative predictions for several field variables from remote sensing data; the best linear regression (r2 = 0.64) for predicting tree canopy cover in Q. ilex stands included Landsat TM channel 7 (TM7). Pereira et al. (1995) used Landsat TM imagery and fieldwork measurements to determine biomass, percent canopy cover and canopy volume in Mediterranean shrublands; the best results were obtained with the Normalized Difference Vegetation Index (NDVI) to estimate percent canopy cover (r2 = 0.65). Oliveira (1998) used field radiometry and Landsat TM imagery to estimate several shrub parameters in the Alto-Daõ e Lafo˜es region (Portugal); field radiometry results showed that the Landsat TM3 channel was the one better correlated with canopy cover (r = 0.91) and the NDVI proved to be the vegetation index (VI) with better performance (r = 0.75); regarding the Landsat TM imagery, it was found that TM3, channel 5 (TM5), and TM7 yielded the better correlation with canopy cover (r = 0.87) and the Atmospherically Resistant Vegetation Index (ARVI) was the best VI (r = 0.86). Calvaõ and Palmeirim (2004) collected field data over Mediterranean shrublands and developed correlations between several biophysical parameters and spectral variables (single channel reflectance and NDVI) from Landsat TM data; the higher correlation for canopy cover was obtained with TM3 (r = 0.91) and NDVI (r = 0.91). 2. Study area and methods The study area is located near the city of E´vora (University of E´vora/Mitra Campus, 388320 N and 88000 W), comprising an area equivalent to a Landsat TM mini-scene (approximately 50 km 50 km). The data used in this study came from two distinct sources: 1-m spatial resolution ortho-rectified digital infrared aerial photography from a flight in the summer of 1995 (acquired between 28 August and 13 September) (Table 1) and 30-m Landsat TM imagery from the same year (acquired on 15 August). Proximity between the dates of the two data sources is

J.M.B. Carreiras et al. / Forest Ecology and Management 223 (2006) 45–53 Table 1 Characteristics of the aerial photos used in this study Aerial photograph #

Tree density

Acquisition date

3535 3643 3685 5153 5195

Low Low High Low High

28 28 28 13 13

August 1995 August 1995 August 1995 September 1995 September 1995

important in this study area because main phenological changes in land cover are highly detectable, due to low tree density. This is more evident in croplands and pastures, which occupy a large proportion of the understorey of Q. suber and Q. ilex ssp. rotundifolia in this region (Gaspar, 1993). At the time of the year of image acquisition (late summer) most of the croplands had been harvested and the pastures were dry. This is a fundamental issue because it allows a better spectral contrast between the overstorey and the understorey. The five aerial photos were selected to cover a broad range of tree densities. Landsat TM mini-scene was calibrated to radiance and apparent reflectance using the time-dependent coefficients proposed by Teillet and Fedosejevs (1995). Image georeferencing was performed with a linear polynomial function, using 21 control points extracted from 1:25 000 scale topographic maps from the Portuguese Army Geographic Institute. A residual mean square error of 9.6 m was obtained. Subsequently, the Landsat TM mini-scene and the digital aerial photos were co-registered, to guarantee spatial correspondence between the two data sources. The methodology used comprised three stages: (a) analysis of digital aerial photos to estimate tree canopy cover; (b) calculation of spectral measures of vegetation cover, derived from satellite imagery; (c) regression modelling of tree canopy cover (aerial photos), as a function of spectral reflectance data (satellite). 2.1. Analysis of digital aerial photos Estimation of tree canopy cover using the aerial photos was accomplished as follows: (i) 30 points were randomly chosen in

47

each digital aerial photo; this value was considered adequate to characterise the spatial distribution of tree canopy cover; (ii) square plots of 120 m 120 m were delineated, centred in the points previously chosen; this plot size was chosen considering the spatial resolution of the satellite imagery, i.e. 4 4 Landsat TM pixels; (iii) in each plot a systematic grid of 196 points was overlaid (14 14); (iv) the number of points falling in tree crowns was counted in each plot (both cork and holm oaks); tree canopy cover (TCC) in each plot was obtained dividing the number of points overlaid in the crowns by the total in the plot, according to TCC = (number of tree crown points/total number grid points) 100%. Fig. 1 displays a plot overlaid on an aerial photo, and the grid used to determine tree canopy cover. It shows the points that were considered to fall on a tree crown (white plus signs) and those falling on the understorey (black circles). 2.2. Satellite image analysis Several spectral measures related to the presence of vegetation were calculated. Vegetation spectral indices (VIs) combine information from two or more spectral channels to enhance vegetation signal, while minimising soil, atmospheric, and solar irradiance effects (Jackson and Huete, 1991). Several VIs were calculated, based on Landsat TM individual channels, namely the NDVI (Rouse et al., 1974), the Soil Adjusted Vegetation Index (SAVI) (Huete, 1988), the modified soil adjusted vegetation index (MSAVI) (Qi et al., 1994), the Green Normalised Difference Vegetation Index (GNDVI) (Gitelson et al., 1996), the ARVI (Kaufman and Tanre´, 1992), the VI5 (Marchetti and Ricotta, 1993), and the VI7 (Lo´pez and Caselles, 1991). The equations for these indices are shown in Table 2. The NDVI, GNDVI, VI5 and VI7 all are normalized difference VIs, and they differ only in the channels used. Classical VIs (e.g., the NDVI) exploit the fact that vegetation is highly reflective in the near infrared (NIR) region and strongly absorbing in the visible (due to chlorophyll absorption) and short wave infrared (SWIR) (due to water absorption).

Fig. 1. (a) Plot insertion in the aerial photo, and (b) detail of the same plot and respective grid (aerial photo 3685, plot 10). The white plus signs represent the grid points overlaid in tree crowns, and the black circles those overlaid in the understorey.

48


Table 2 Spectral indices calculated with the Landsat data. TM1, TM2, TM3, TM4, TM5, TM7 – reflectances of Landsat TM channels. L: soil adjustment factor; a value of 0.5 is reasonable for a broad range of soil types (Huete, 1988). g: atmospheric self-correcting factor, dependent on aerosols type; when the canopy is sparse and atmospheric data are unknown, a value of g = 1 is recommended for most remote sensing applications (Kaufman and Tanre´, 1992) Spectral indices/K–T components

Equation

NDVI SAVI MSAVI GNDVI ARVI VI5 VI7 K–T brightness K–T greenness K–T wetness

(TM4 TM3)/(TM4 + TM3) ((TM4 TM3)/(TM4 + TM3))(1 + L) (2TM4 + 1 ((2TM4 + 1)2 8(TM4 TM3))1/2)/2 (TM4 TM2)/(TM4 + TM2) (TM4 RB)/(TM4 + RB), RB = TM3 g(TM1 TM3) (TM4 TM5)/(TM4 + TM5) (TM4 TM7)/(TM4 + TM7) 0.2043TM1 + 0.4158TM2 + 0.5524TM3 + 0.5741TM4 + 0.3124TM5 + 0.2303TM7 0.1603TM1 0.2819TM2 0.4934TM3 + 0.7940TM4-0.0002TM5 0.1446TM7 0.0315TM1 + 0.2021TM2 + 0.3102TM3 + 0.1594TM4 0.6806TM5 0.6109TM7

This spectral behaviour contrasts with that of most other land surfaces and atmospheric features and is the key to spectral detection of vegetation and quantification of its abundance. The VI5 and VI7 replace the visible channel of classical vegetation indices with SWIR channels. These are sensitive to vegetation moisture, enhancing the contrast between vegetation and drier areas, and are less affected by atmospheric effects. The GNDVI uses the green spectral region instead of the red region, to increase sensitivity to the presence of chlorophyll (Gitelson et al., 1996). Huete (1988) developed a VI (SAVI) to minimise confounding effects induced by the soil background, especially in areas with relatively sparse vegetation cover. Qi et al. (1994) proposed a modification of SAVI (MSAVI); it uses a variable soil adjustment factor that depends on the amount of vegetation present. Kaufman and Tanre´ (1992) developed the ARVI, meant to be resistant toward atmospheric effects; it uses the blue channel to quantify, and compensate for, the magnitude of the atmospheric effect. Other spectral indicators of the presence of vegetation were also calculated, namely the components of the Kauth–Thomas (K–T) transformation, also know as the tasselled cap transformation (Crist and Cicone, 1984). The K–T transformation converts the six reflective channels of the Landsat TM into three orthogonal components, which are linear combinations of the original channels. The first of these component, designated brightness, expresses differences in soil properties, like particle size and organic matter content; the second component, greenness, is well correlated with tree canopy cover, leaf area index and live biomass; the third component, wetness, is sensitive to soil and plant moisture (Crist and Cicone, 1984). The Landsat TM satellite data were converted to apparent reflectance, so we used the K–T equivalent transformation for reflectance factor data (Crist, 1985).

spectral variables. Simple linear regression models were developed to estimate tree canopy cover as a function of each VI. Multiple linear regression was used for the models based on the components of the K–T transformation and for the model based on individual Landsat TM channels. For the latter model, variables were chosen using a stepwise forward selection procedure, with penter = 0.05 and premove = 0.40. Model evaluation was performed using goodness-of-fit statistics (coefficient of determination (r2) and adjusted r2). Model validation was based on the analysis of the difference between observed and predicted values (residuals) for the validation subset (n = 36), using some of the tests for model assessment recommended by Soares and Tome´ (1993). Residual analysis was performed using the mean prediction residuals, mean absolute prediction residuals, and sum of squared residuals. The statistical tests were performed to evaluate model bias. First, the Shapiro-Wilk W test (Shapiro et al., 1968) was used to test for normality of the distribution of prediction residuals. Student’s t-test and regression tests were performed for models with normally distributed residuals. Student’s t-test evaluates model accuracy, based on the null hypothesis of a mean prediction residual equal to zero. The regression tests are based on the null hypothesis that when observed values are regressed over the corresponding predicted values, a regression line with unit slope and zero intercept should be obtained, if the model is unbiased. For models with non-normally distributed residuals the non-parametric Wilcoxon’s signed-rank test was used. This test is based on the sign and magnitude of the rank of the differences between pairs of measurements. The formal null hypothesis for Wilcoxon’s test is that the population distribution of differences is symmetrical about zero. Details about all these tests can be found in standard statistics textbooks (e.g., Conover, 1980; Steel and Torrie, 1980; Zar, 1999).

2.3. Model adjustment and validation 3. Results The 142 tree canopy cover estimates were randomly split into two subsets, one for model fitting, with 75% of the data (n = 106), and another for validation, with the remainder 25% (n = 36). Ordinary least squares linear regression modelling was employed to estimate tree canopy cover as a function of

3.1. Determination of tree canopy cover Table 3 shows a summary of the results of the estimation of tree canopy cover in the five aerial photos. Some of the aerial

J.M.B. Carreiras et al. / Forest Ecology and Management 223 (2006) 45–53 Table 3 Descriptive statistics of tree canopy cover estimation in the aerial photos Aerial photograph # 3535 3643 3685 5153 5195

Number of plots 30 30 27 27 28

Tree canopy cover statistics Mean

Minimum

Maximum

S.D.

31.6 7.4 27.5 11.6 28.5

2.0 0.0 0.0 0.0 0.0

61.2 42.3 56.1 48.2 100.0

15.2 13.0 16.7 15.4 30.6

49

Table 4 Performance of the nine tested models, showing coefficients estimates, coeffi2 cient of determination (r2) and adjusted r2 (radj ) (n = 106); TCC: tree canopy cover; BRIGHT, GREEN, WET – K–T brightness, greenness, and wetness components, respectively Model

r2

2 radj

TCC = 36.298 + 278.354NDVI TCC = 44.472 + 552.388SAVI TCC = 42.502 + 628.068MSAVI TCC = 87.977 + 397.043GNDVI TCC = 12.491 + 152.326ARVI TCC = 57.223 + 231.738VI5 TCC = 8.569 + 182.749VI7 TCC = 54.723 36.617BRIGHT + 746.741GREEN + 157.412WET 9. TCC = 63.626 447.222TM5 + 623.837TM4 714.626TM3 + 281.354TM7

0.72 0.66 0.64 0.64 0.69 0.59 0.63 0.70

0.71 0.66 0.64 0.64 0.69 0.59 0.63 0.69

0.74

0.73

1. 2. 3. 4. 5. 6. 7. 8.

data is concentrated in the classes with lower canopy cover, which is to be expected in these relatively open woodlands. 3.2. Model adjustment and validation

Fig. 2. Histogram of tree canopy cover (%) for all the plots (n = 142).

photos have fewer than 30 plots because the corresponding grid point was not totally inside the photo, so those plots were rejected. Fig. 2 shows the histogram for tree canopy cover, considering all plots (n = 142). As expected, the majority of the

The results of the adjustment of the different models can be seen in Table 4. The model with the best performance is the one resulting from stepwise regression (with individual channels as independent variables), followed by the model with the NDVI as independent variable, and then those using the three K–T components or the ARVI. To get more insight into the data, a 2D perspective, i.e. displaying several independent variables versus the dependent variable, is presented. In Fig. 3, the different scatterplots, using

Fig. 3. Scatterplots of tree canopy cover (n = 142) vs. Landsat TM individual channels: (a) TM3; (b) TM4; (c) TM5; (d) TM7.

50


the individual Landsat TM channels are represented, and in Fig. 4 the same representation is shown for the NDVI, ARVI, and K–T components. Regarding Fig. 3, there is a non-linear inverse relation with tree canopy cover, with the exception of TM4 reflectance, where there is no correlation. In Fig. 4, as would be expected for the VIs, the highest tree canopy cover corresponds to the highest VIs values. Moreover, the highest K– T brightness component occurs in those plots with lowest tree canopy cover; the opposite happens for the K–T greenness and wetness components. Table 5 shows model validation results. The models with lower mean prediction residuals (rp) (models 2, 3, and 7) have, however, higher mean absolute prediction residuals (arp) and sum of squared residuals (ssr). This could result from both higher negative and positive residuals that sum to a lower value (lower rp) but resulting in higher arp and ssr; therefore the arp and ssr measures seem to be more appropriate to assess model

validation. The four models with better goodness-of-fit are also those with lower arp and ssr; model 9 (with Landsat channels TM3, TM4, TM5 and TM7) has the lowest arp and ssr, followed by model 8 (with K–T components), model 1 (with NDVI), and model 5 (with ARVI). The W value of the Shapiro–Wilk test is significant (for a = 0.05) for models 1–3, 6, and 7. The results of the t-test indicate that there is insufficient information to reject the null hypothesis for all the tested models (for a = 0.05). The statistic of the intercept = 0 regression t-test is significant for model 3 (for a = 0.05), and for models 6 and 7 (for a = 0.001). The slope = 1 regression t-test is significant for models 1, 2, and 9 (for a = 0.05), for model 3 (for a = 0.01), and for models 6 and 7 (for a = 0.001). These results indicate that models 6 and 7 are highly biased (for a = 0.001), and that models 1, 2, 9, and 3 are slightly biased. However, we should note that the previous t-tests only have meaningful interpretation for those models displaying normally distributed residuals.

Fig. 4. Scatterplots of tree canopy cover (n = 142) vs. NDVI, ARVI, and K–T components: (a) NDVI; (b) ARVI; (c) K–T brightness; (d) K–T greenness; (e) K–T wetness.


51

Table 5 Summary statistics of prediction residuals and results of statistical tests of model validation (n = 36); rp: mean prediction residuals, arp: mean absolute prediction residuals, ssr: sum of squared residuals; the Wilcoxon’s signed-rank test is not shown for models 4, 5, 8, and 9, as there is no evidence to reject the null hypothesis of the Shapiro–Wilk test (for a a = 0.05) Model

Summary statistics rp

1 2 3 4 5 6 7 8 9

0.66 0.38 0.52 0.79 1.83 0.67 0.64 0.91 0.73

arp

7.82 8.76 9.21 8.77 7.90 24.43 24.28 7.20 6.59

Statistical tests ssr

3717.43 4448.15 4967.28 4304.25 3374.09 41117.04 41317.05 2988.73 2713.60

Shapiro–Wilk test (W)

0.930 0.938 0.937 0.973 0.990 0.902 0.900 0.982 0.953

The Wilcoxon z-test, appropriate for residuals with a nonnormal distribution, indicates that there is not enough information to reject the null hypothesis for the tested models (for a = 0.05). Therefore, the previously mentioned four models with better goodness-of-fit, and lower prediction residuals statistics, also passed on the appropriate statistical tests, the exception being model 9 for the slope = 1 regression ttest. Fig. 5 shows the scatterplots of observed versus predicted tree canopy cover (in the validation subset) for these four models. Models 8 and 9 yielded a better agreement between observed and predicted data. Another relevant result in these models is the tree canopy cover overestimation for values

t-test (t)

0.385 0.201 0.261 0.427 1.118 0.118 0.112 0.591 0.498

Regression t-test

Wilcoxon test (z)

Intercept = 0

Slope = 1

1.442 2.002 2.155 1.606 0.292 4.945 4.822 0.989 1.512

2.399 2.639 2.812 1.813 1.352 6.794 6.814 1.917 2.638

0.220 0.377 0.518 –

– 0.896 0.676 – –

greater than approximately 60%, although this could be due to an insufficient number of observations in this range. 4. Discussion At this time of the year (late summer) the understorey has a very light colour, which is due to the presence of dry vegetation and soil. The tree canopy is darker and when tree canopy cover increases the reflectance of the plot decreases (reflective TM3), because the soil reflects solar radiation and the chlorophyll absorbs it (Fig. 3a). In the middle infrared (TM5 and TM7) the situation is the same, but the absorption is due to the presence of water in the living tissues (Fig. 3c and d). In the NIR (TM4)

Fig. 5. Observed vs. predicted tree canopy cover (%) for the (a) model 1, (b) model 5, (c) model 8, and (d) model 9. The data refers to the validation subset (n = 36). The 1:1 line represents perfect agreement.

52


crowns are highly reflective, as well as the understorey of bright soils and dry grass, thus resulting in the weak correlation shown in the scatterplot (Fig. 3b). Satellite imagery acquired in the late summer maximize the spectral contrast between the evergreen tree crowns and the dry herbaceous background. However, the study area is still characterized by an extremely variable understorey, containing bare soil, dry grass, and a few evergreen shrubs. The differences between VIs values of plots with the same tree canopy cover are due to the presence of different types of understorey, which have different spectral properties, and/or different vegetation structure. Huete (1989) refers that in areas with significant soilbrightness changes that arise from moisture differences, roughness variations, shadow, or organic-matter differences, there are soil-induced influences on the vegetation indices values, and that these effects are predominant in partially vegetated canopies, as in this study. To distinguish different plots according to tree canopy cover, a diagram representing Landsat channels TM3 and TM4 was built (Fig. 6). The plots were grouped into five classes, according to tree canopy cover estimated from the aerial photos. Richardson and Wiegand (1977) observed that bare soil pixels were generally aligned according to a straight line (beginning in the origin), which approximately bisects the redNIR space (soil line). They also observed that points representing pixels of water were located close to the origin of the soil line, while pixels of vegetation were always located to the left of the same line. It is evident that plots with lower tree canopy cover (i.e. with higher proportion of understorey) are located closer to the soil line. The higher or lower reflectance of these plots in these channels is related with the type of the background. An increase in tree canopy cover, within the same type of understorey, shows an increase of the reflectance in the NIR (TM4) and a decrease in the red (TM3). Considering the four best models, model 9 is the harder to interpret, because it is a multiple linear regression model. Those with the NDVI (model 1) or with the ARVI (model 5) are the

simplest, followed by that with the K–T components. All the parameters in model 9 are interpretable, except the one associated with the TM7; which should be negative, like the one associated with TM5, reflecting the presence of water in the living tissues, typical of the SWIR spectral region; one possible explanation is the observed high correlation between TM5 and TM7 (r = 0.96); the signal is positive for TM4 because a higher tree canopy cover is associated with a high value of reflectance in the NIR; the opposite happens with TM3, the negative signal indicating that a higher absorption (less reflectance) is a sign of the presence of high values of chlorophyll, and so more vegetation. As expected, the equation with the NDVI or the ARVI present the parameter with a positive signal, indicating a direct relation between the increase of these VIs and tree canopy cover. The equation with the K–T components has the parameters with the appropriate signals: areas with dense canopy cover are dark (negative brightness coefficient), green (positive greenness sign), and contain abundant leaf tissue moisture (positive wetness sign). 5. Conclusions The low tree density characteristic of these Mediterranean evergreen oak woodlands has a large influence on the estimation of tree canopy cover with remote sensing data. The spectral contribution of the understorey for the signal received by the satellite sensor increases with decreasing tree density. Therefore, one should select a period when there is maximum spectral contrast between the overstorey and the understorey. The model with best predictive capability (model 9) includes one channel in the visible domain, one in the NIR, and two in the SWIR, and produces good estimates of tree canopy cover. The second best model (model 8) includes the components of the K–T transformation, yielding a robust and easily interpretable model. Two other models present good predictive capability, namely those including the NDVI (model 1) and the ARVI (model 5). It is also important to stress that the type of relation established can be helpful in understanding the dynamics of tree canopy cover in this region of the country. The use of high and medium spatial resolution satellite imagery could be an instrument to monitor changes that occur in this type of forest, with lower tree density. Another important key aspect is the presence of several types of understory associated with different crown cover percentage, which favours the robustness of the model being built. Acknowledgements

Fig. 6. Representation of the dataset (142 plots), based in tree canopy cover (TCC) classes, in the TM3–TM4 reflectance plane.

The authors gratefully acknowledge the two anonymous reviewers for their helpful comments and suggestions. This study was funded by the Portuguese ‘‘Subprograma Cieˆncia e Tecnologia do 28 Quadro Comunita´rio de Apoio’’ within the grant Ref. GGP XXI/BTI/4458/96. This grant was integrated in a broader project called ‘‘Monitorizaçaõ da Resposta ao Stress Ambiental em Povoamentos de Sobreiro’’, Ref. PRAXIS XXI/ 3/3.2/FLOR/2122/95.


References Baldocchi, D.D., Xu, L., Kiang, N., 2004. How plant functional type, weather, seasonal drought, and soil physical properties alter water and energy fluxes of an oak-grass savanna and an annual grassland. Agric. For. Meteorol. 123, 13–39. Blanco, E., Casado, M.A., Costa, M., Escribano, R., Garcia, M., Genova, M., Gomez, A., Go´mez, F., Moreno, J.C., Morla, C., Regato, P., Sainz, H., 1997. Los Bosques Ibe´ricos: Una Interpretacioń Geobotaˆnica. Editorial Planeta, Barcelona, Spain. Brasier, C.M., 1992. Oak tree mortality in Iberia. Nature 360, 539. Brasier, C.M., 1996. Phytophthora cinnamomi and oak decline in southern Europe Environmental constraints including climate change. Annales des Sciences Forestie`res 53, 347–358. Calvaõ, T., Palmeirim, J.M., 2004. Mapping Mediterranean scrub with satellite imagery: biomass estimation and spectral behaviour. Int. J. Remote Sens. 25 (16), 3113–3126. Conover, W.J., 1980. Practical Non-parametric Statistics, 2nd ed. John Wiley and Sons, New York. Crist, E.P., 1985. A TM tasselled cap equivalent transformation for reflectance factor data. Remote Sens. Environ. 17, 301–306. Crist, E.P., Cicone, R.C., 1984. A physically-based transformation of thematic mapper data – the TM tasseled cap. IEEE Trans. Geosci. Remote Sens. 22 (3), 256–263. David, T.S., Ferreira, M.I., Cohen, S., Pereira, J.S., David, J.S., 2004. Constraints on transpiration from an evergreen oak tree in southern Portugal. Agric. For. Meteorol. 122, 193–205. DGF, 2001. Inventa´rio Florestal Nacional. Portugal Continental. 3 Revisaõ, 1995–1998. Relato´rio Final. Direcçaõ-Geral das Florestas, Lisbon, Portugal. Gaspar, J., 1993. As Regio˜es Portuguesas. Ministe´rio do Planeamento e da Administraçaõ do Territo´rio, Secretaria de Estado do Planeamento e Desenvolvimento Regional, Lisbon, Portugal. Gitelson, A.A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sens. Environ. 58, 289–298. Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25, 295–309. Huete, A.R., 1989. Soil influences in remotely sensed vegetation-canopy spectra. In: Asrar, G. (Ed.), Theory and Applications of Optical Remote Sensing. John Wiley & Sons, New York, pp. 107–141. Jackson, R.D., Huete, A.R., 1991. Interpreting vegetation indices. Prev. Vet. Med. 11, 185–200. Joffre, R., Lacaze, B., 1993. Estimating tree density in oak savanna-like dehesa of southern Spain from SPOT data. Int. J. Remote Sens. 14 (4), 685–697. Joffre, R., Rambal, S., Damesin, C., 1999. Functional attributes in Mediterranean-type ecosystems. In: Pugnaire, F.I., Valladares, F. (Eds.), Handbook of Functional Plant Ecology. Marcel Dekker, New York, pp. 347–380. Kaufman, Y.J., Tanre´, D., 1992. Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Trans. Geosci. Remote Sens. 30 (2), 261– 270.

53

Lo´pez, G.M.J., Caselles, V., 1991. Mapping burns and natural reforestation using thematic mapper data. Geocarto Int. 1, 31–37. Lourenço, N., Pinto-Correia, T., Jorge, M.R., Machado, C.R., 1998. Farming strategies and land use changes in southern Portugal: land abandonment or extensification of the traditional systems? Mediterraˆneo 12/13, 191– 208. Marchetti, M., Ricotta, C., 1993. L’impiego di dati LANDSAT TM per il monitoraggio della ripresa vegetativa in aree incendiate. Monti e Boschi 3, 22–26. Oliveira, T.M., 1998. Cartografia Quantitativa de Formaço˜es Arbustivas Empregando Dados de Detecçaõ Remota. Tese de Mestrado em Gestaõ dos Recursos Naturais. Universidade Tećnica de Lisboa, Instituto Superior de Agronomia, Lisbon, Portugal. Pereira, J.M.C., Oliveira, T.M., Paul, J.C.P., 1995. Satellite-based estimation of Mediterranean shrubland structural parameters. EARSeL Adv. Remote Sens. 4 (3), 14–20. Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A modified soil adjusted vegetation index. Remote Sens. Environ. 47, 1–25. Richardson, A.J., Wiegand, C.L., 1977. Distinguishing vegetation from soil background information. Photogrammet. Eng. Remote Sens. 43 (12), 115– 125. Romero-Calcerrada, R., Perry, G.L.W., 2004. The role of land abandonment in landscape dynamics in the SPA ‘Encinares del rıó Arberche y Cofio’ Central Spain1984–1999. Landsc. Urban Plan. 66, 217–232. Rouse, J.W., Haas, R.W., Schell, J.A., Deering, D.W., Harlan, J.C., 1974. Monitoring the vernal advancement and retrogradation (greenwave effect) of natural vegetation. NASA/GSFCT Type III Final Report, Greenbelt, MD, USA. Salvador, R., Pons, X., 1998. On the applicability of Landsat TM images to Mediterranean forest inventories. For. Ecol. Manage. 104, 193–208. Shapiro, S.S., Wilk, M.B., Chen, H.J., 1968. A comparative study of various tests of normality. J. Am. Stat. Assoc. 63, 1343–1372. Soares, P., Tome´, M., 1993. Procedures for validation of forest management growth models. An application to continuous forest inventory data. In: Vanclay, K., Skousgaard, J.P., Gertner, G.T. (Eds.), Growth and Yield Estimation from Successive Forest Inventories. Proceedings from IUFRO Conference, Copenhagen. Forskningsserien 3, pp. 167–180. Steel, R.G.D., Torrie, J.H., 1980. Principles and Procedures of Statistics, 2nd ed. McGraw-Hill, Singapore. Stoate, C., Boatman, N.D., Borralho, R.J., Rio Carvalho, C., de Snoo, G.R., E´den, P., 2001. Ecological impacts of arable intensification in Europe. J. Environ. Manage. 63, 337–365. Swain, P.H., Davies, S.M., 1978. Remote Sensing. The Quantitative Approach. McGraw-Hill, New York. Teillet, P.M., Fedosejevs, G., 1995. On the dark target approach to atmospheric correction of remotely sensed data. Can. J. Remote Sens. 21 (4), 374–386. Terradas, J., 1999. Holm oak and holm oak forests: an introduction. In: Roda, F., Retana, J., Gracia, C.A., Bellot, J. (Eds.), Ecology of Mediterranean evergreen oak forests. Springer-Verlag, Berlin, pp. 3–14. Zar, J.H., 1999. Biostatistical Analysis, 4th ed. Prentice-Hall, Upper Saddle River, NJ, USA.