Forest Ecology and Management 259 (2010) 12151224

Contents lists available at ScienceDirect

Forest Ecology and Management

journal homepage: www.elsevier.com/locate/foreco

Sensitivity of landscape pattern metrics to classication approaches

Jean-Franc ois Mas *, Yan Gao, Jose Antonio Navarrete Pacheco
tzcuaro No. 8701, Col. Ex-Hacienda de San Jose de La noma de Me xico (UNAM), Antigua Carretera a Pa Centro de Investigaciones en Geografa Ambiental, Universidad Nacional Auto n, Huerta, C.P. 58190 Morelia, Michoaca Mexico

A R T I C L E I N F O

A B S T R A C T

Article history: Received 7 February 2009 Received in revised form 26 October 2009 Accepted 18 December 2009 Keywords: Landscape patterns metrics Fragmentation indices Object-oriented classication Pixel-oriented classication

Computing of landscape pattern metrics from spectrally classied digital images is becoming increasingly common. Recently, object-orientated image classication is being seen as an alternative and is tending to replace pixel-based approaches. However, object-based methods are likely to inuence and produce biases in the results of these spatial analyses. In this study, the sensitivity of 85 landscape metrics to different classication methods and parameters are analyzed. A Landsat image of a complex mountainous forest region of Mexico was classied using pixel-based and object-based approaches. Nine object-based classied images were obtained using a region-growing algorithm based upon different segmentation parameters. Pixel-based classied images were smoothed using different methods (majority ltering, sieving and clumping). Accuracy assessment was carried such that classied images with similar accuracy were compared. Landscape metrics were then derived from the different classied images and compared through a coefcient of variation computing. Almost all the metrics showed variability due to classication and post-processing methods, particularly core area metrics and some proximity and contagion/interspersion indices. Caution must be observed when comparing values of metrics derived from images with slight differences in their characteristics or in the way they have been processed as, for example, in landscape monitoring studies based upon multidate imagery. 2010 Elsevier B.V. All rights reserved.

1. Introduction Numerous studies have shown that the spatial patterns of landscape may have signicant inuences on ecological processes, such as population dynamics, biogeochemical cycling, and biodiversity. Therefore, identifying and characterizing spatial patterns of landscape are often necessary in landscape ecological studies. Over the last decades, such studies have beneted from a proliferation of metrics for characterising landscapes (Frohn, 1998; Gustafson, 1998; McGarigal and Cushman, 2002; Vogt et al., 2007). The digital nature of land cover information from satellite imagery enables a potentially large number of landscape metrics to be derived (Haines-Young and Chopping, 1996). The large area coverage and repeat viewing of remotely sensed data provides information over a considerable range of spatial and temporal resolution for mapping land cover. Typically, remote sensing images are classied using statistical (maximum likelihood) algorithms that utilise the spectral values of individual pixels. These methods do not make use of spatial information in the image: each pixel is classied independently taking into account only its spectral response and ignoring the response of the neighbouring pixels. As a consequence, the resulting thematic

* Corresponding author. Tel.: +52 443 322 38 35; fax: +52 443 322 38 80. E-mail address: jfmas@ciga.unam.mx (J.-F. Mas). 0378-1127/$ see front matter 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.foreco.2009.12.016

maps usually suffer from a salt-and-pepper effect which is the presence of isolated pixels that belong to a different class than their neighbouring pixels. Due to the limitations of pixel-based methods, alternative approaches, such as contextual and objectoriented classication, which use neighbouring pixel information, are arising. The development of these methods stems primarily from the desire to use the important semantic information necessary to interpret an image, which is not presented in single pixels but rather in meaningful objects and their mutual relations. In object-oriented classication, homogeneous image objects are rst extracted and subsequently classied. The mapping results thus represent real-world objects and lack the salt-and-pepper appearance of pixel-based classied images. An overview of these approaches can be found in Walter (2004). Object-oriented approaches are increasingly used by the remote sensing community, because generally they lead to more accurate maps. The characteristics of the remotely sensed images and the methodologies applied for their processing and classication may strongly inuence the spatial characteristics of the land cover data from which spatial metrics are calculated. For instance, many researchers have investigated the inuence of spatial resolution (also referred to as grain in the literature) on landscape metrics. These authors found that the effects depend upon the metrics under consideration, and that there exists a spatial resolution beyond which the resolution effects are no longer obvious (Turner et al., 1989; Benson and MacKenzie, 1995; Riitters et al., 1995; Qi

1216

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

Fig. 1. Location of the study area. Left side of the gure are two sketch maps indicating Mexico and Michoacan state where the study area is located; right side is the false colour composite of Landsat image.

and Wu, 1996). Attention has also been devoted to the effect of the extent of the spatial area over which pattern metrics are estimated. Saura (2002) analyzed the effects of the minimum mapping unit on several commonly used landscape metrics and found that a larger minimum mapping unit implies the underestimation of landscape diversity and fragmentation. Kearns et al. (2005) analyzed the redundancy of metrics and the sensitivity to changes in extent. Brown et al. (2000) tested the effects of the amount of forest cover, the phenology, the atmospheric variability and the post-classication processing on the consistency of four metrics. They found that some metrics (average patch size and number of patches) were more sensitive than others (proportion of cover and edge density). Baldwin et al. (2004) examined the sensitivity of landscape metrics to spatial extent, spatial resolution and thematic resolution. Recently, new processing approaches are being developed to classify remotely sensed data, such as articial neural networks, object-oriented classication, fuzzy classication, and others. Pixel-based and object-based oriented approaches produce classied images with different spatial patterns. However, there is no study aimed at assessing the effects of these classication approaches in landscape pattern evaluation. This paper aims at assessing the effects of the classication method (pixel-based versus object-based) and of the parameters used to carry out object-oriented classications on landscape metrics. A supercial analysis suggests that pixel-based classications lead to more fragmented classied images, and that indices which are sensitive to the presence of small patches are more sensitive to the image processing approach. 2. Study area and data The study area is located in the State of Michoacan, central west Mexico, within longitude 1028000 W and 1028320 W, and latitude 198020 N and 198360 N, and covers an area of approximately 58 km 60 km (Fig. 1). It is a mountainous region, with elevation ranging from 220 m to 3830 m. The area is a complex mosaic of several land cover types including temperate pine and oak forest, dry tropical forest, orchard (mainly avocado plantations), bare soil (lava ow), crops and pasture lands (including secondary dry tropical forest used as pasture). This spatially complex area was

chosen to highlight the differences between the image processing approaches. 3. Material The available data comprise a Landsat ETM+ image obtained on 16/Feb/2003; ortho-corrected air photographs and a land cover map from the National Forest Inventory 2000. A geometric correction of the image was previously carried out using 86 ground control points extracted from the ortho-corrected photographs with a RMS error below one pixel (16.5 m). Image segmentation was performed using the image processing package SPRING (Camara et al., 1996), which is a non-commercial programme ranked second in segmentation quality among seven algorithms tested by Meinel and Neubert (2004). The landscape metrics were generated using the FRAGSTATS program version 3.3 (McGarigal and Cushman, 2002; McGarigal et al., 2002). 4. Methods The research had several major components: (i) land use/cover classication; (ii) accuracy assessment of classied images; (iii) computation of landscape metrics; (iv) sensitivity measurement; and (v) interpretation and analysis. This section provides the technical details for the rst four procedures (Fig. 2). 4.1. Image classications Training areas for eight land cover categories (irrigated agriculture, rainfed agriculture, grasslands, orchards, dry tropical forest, temperature forest, human settlement and, bare land) were dened using a colour composite of the Landsat image, the air photographs and eld data. Classications were carried out using the standard pixel-based maximum likelihood method and an object-oriented classication, based upon two steps. First a segmentation was carried out. This consisted of grouping neighbouring pixels with a similar spectral response into the same object. The segmentation was based upon a region-growing algorithm: segments are formed starting from suitable initial pixels (seeds) by iteratively augmenting them with neighbouring pixels that satisfy a chosen homogeneity criteria. In the second

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

1217

Fig. 2. Flow chart of image processing steps.

step, the entire objects (segments) were classied by majority voting whereby pixels of each object are assigned to the most frequent class using the pixel-based. In order to control the segmentation procedure, two parameters were used, similarity and area. Similarity is a threshold value that determines whether two neighbouring objects are merged, while the area threshold is used to lter out the objects smaller than this value (control of the minimum mapping unit) (Bins et al., 1996). A set of classied images was produced using the same parameter area and a range of values for similarity.

Pixel-based classied images present a speckled appearance due to the salt-and-pepper effect. To suppress this effect, postclassication processing techniques were applied so that regions with less than a minimal area were removed. Image processing packages provide three common methods for generalizing a classied image and remove the smallest patches: (i) majority ltering, which changes the class of every pixel to the dominant class found within a specied search window around each pixel; (ii) sieving, which looks at the neighbouring 4 or 8 pixels to determine if a pixel is grouped with the pixels of the same class. If the number of pixels that are grouped are less than the threshold

Fig. 3. Performance of the three post-classication methods with various thresholds. These methods allow eliminating small groups of pixels and present different results.

1218 Table 1 Confusion matrix used in the Mc Nemar test. Method 2 Method 1 Correct Correct Incorrect f11 f12

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

applied to eliminate small patches by recoding to the class with which they share their longest border.
Incorrect f21 f22

4.2. Classication accuracy The accuracy of classication for all images was evaluated with ground data comprising 305 random points. Points category was determined by photo-interpretation of the air photographs and by eld visits for ambiguous cases. The accuracy values of the classied images were compared to evaluate the signicance of the difference in accuracies of each pair of classications. The same set of ground truth data were used to assess the accuracy of the classied images to be compared, that is the samples are related. Therefore, the statistical signicance of the difference between the two accuracy statements was evaluated using McNemars test, because this test takes into account the lack of independence between samples. It is a non-parametric test based on confusion matrixes that are 2 2 in dimension (Table 1). Attention is focused on the binary distinction between correct and incorrect class allocations. The McNemars test is based on the standardized normal test statistic (Foody, 2004). f 12 f 21 Z p f 12 f 21 (1)

value, these pixels are removed from the class and no data pixels are left; (iii) clumping, which groups similar classied areas together by rst performing a dilate operation and then an erode operation on a classied image using the specied kernel size. The performances of those three post-classication methods are illustrated in Fig. 3. Two approaches were used in this work in order to remove the speckled appearance: (1) applying a clumping (3 3 pixels) analysis to remove the scattered pixels; (2) applying a majority lter (3 3 pixels), and then sieving to remove the remaining small patches, in which two sieving threshold values were applied, based on kernels of 2 2 pixels and 5 5 pixels. These two processes produced three pixel-based classied images. In order to eliminate patches with an area below the minimum threshold area used during the segmentation procedure, a GIS procedure was

Table 2 List of landscape metrics at the class and landscape levels, respectively. Some indices, such diversity indices can only be calculated at landscape level. Other indices are not calculated at landscape level because they have no meaning (e.g. percentage of landscape) or are correlated totally with another index (e.g. number of patches and patch density). Indices based on averaging values obtained for each patch were calculated by simple mean (annotated_MN) and area-weighted mean (annotated_MA). Structural feature Area/density/edge Index (acronym) NP PLAND PD LPI TE LSI AREA_MN, _AM GYRATE_MN, _AM SHAPE_MN, _AM FRAC_MN, _AM PARA_MN, _AM CONTIG_MN, _AM PAFRAC TCA CPLAND NDCA CORE_MN, _AM DCORE_MN, _AM CAI_MN, _AM PROX_MN, _AM ENN_MN, _AM CLUMPY CONTAG PLADJ IJI DIVISION MESH SPLIT AI CONNECT COHESION PR SHDI SIDI MSIDI SHEI SIEI MSIEI Full name (unit) Number of parches Percentage of landscape (%) Patch density (#/100 ha) Largest Patch Index (%) Total edge (m) Lanscape shape index Mean patch area (ha) Mean radius of gyration (m) Mean shape index Mean fractal dimension index Mean perimeterarea ratio Mean contiguity index Perimeterarea fractal dimension Total core area (ha) Core area percent of landscape (%) Number of disjunct core areas Mean core area (ha) Mean disjunct core area (ha) Mean core area index (%) Mean proximity index (m) Mean Euclidian nearest neighbour distance (m) Clumpiness Contagion (%) Proportion of like adjacencies (%) Interspersion/juxtaposition index (%) Landscape division index (%) Effective mesh size (ha) Spitting index Aggregation index (%) Conectance index (%) Patch cohesion index Patch richness Shannons Diversity Index Simpsons Diversity Index Modied Simpsons Diversity Index Shannons evenness index Simpsons evenness index Modied Simpsons evenness index Class level Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes No No No No No No No Landscape level Yes No No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Shape

Core area

Isolation/proximity

Contagion/interspersion

Connectivity

Diversity

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

1219

in which fij indicates the frequency of verication samples lying in confusion matrix element i and j. f12 and f21are the number of pixels that were correctly classied with one method but incorrectly with the second. The square of Z follows a chi-square distribution with one degree of freedom. So the test equation can be expressed as Z2 f 12 f 21 f 12 f 21
2

(2)

total of 45 metrics. At the class level, there are a total of 40 metrics, falling within ve of the six major groups, because diversity metrics are based upon class diversity and therefore do not apply at the class level. A detailed description of the metrics can be found in McGarigal et al. (2002) and Saura (2002). The patch core area was computed at 100 m from the edge. In order to calculate proximity metrics, the search radius was set to 5000 m. These metrics were computed for each of the classied images. 4.4. Sensitivity measurement In order to evaluate the sensitivity of each metric to the method of classication, standard deviations of the values of each metric, as a measure of the dispersion without reference to the direction or bias, were used. The standard deviations are reported in the units of the metric. To standardize the measure for comparison between landscape metrics, standard deviation values were divided by the mean of the metric values to obtain a value referred to as the coefcient of variation. This calculation yields a quantity that summarizes the relative amount of variation, expressing the standard deviation as a proportion of the average metric value. 5. Results 5.1. Image classication and accuracy assessment Nine segmentations were generated with similarity thresholds ranging from 19 to 59 in intervals of 5, and a constant area

with the derived value compared against tabulated chi-square values to indicate its statistical signicance. Only classied images which present a similar accuracy (no signicant difference at p = 0.05) were considered in the following steps. This was done in order to avoid comparing landscape metrics derived from images which present different values of accuracy due to the choice of inadequate parameters during segmentation. 4.3. Computation of landscape metrics A total of 85 metrics (Table 2) were considered in the context of the research objective and landscape ecology principles (Turner et al., 1989; Forman, 1995; McGarigal and Cushman, 2002). These metrics are related to landscape composition (e.g., proportional abundance of each class) or landscape conguration (e.g., patch size distribution and density, patch shape complexity, and interspersion). They are grouped into six major categories: (1) area/density/ edge, (2) shape, (3) core area, (4) isolation/proximity, (5) contagion/ interspersion and (6) diversity. At the landscape level, there are a

Fig. 4. Segmentations with similarity threshold from 19 to 59 with intervals of 5, and a constant area threshold of 22.

1220

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

Table 3 Pixel-based and segment-based images classication overall accuracies (%). Pixel-based classications Clumping Majority ltering + sieving (threshold 2) Majority ltering + sieving (threshold 5) Object-oriented classications Value of similarity parameter during segmentation 19 73.3 75.5 75.5 68.7 24 71.8 29 72.5 34 73.0 39 73.5 44 71.7 49 72.3 54 71.2 59 69.7

threshold of 22. The selection of the similarity parameter was based on visual checking of the segmentation results. With value 19, it appeared that the image was over-segmented, and with value 59 it was under-segmented (Fig. 4). So the tested similarity values were set between these two values. The area threshold set at 22 pixels was in accordance with recommendations by Espindola et al. (2006) who used a Landsat image and found optimal segmentation results with this area threshold. This value corresponds to a minimum mapping unit of 2 ha. The segmented images were classied along with the unsegmented original image (pixel-based classication) and the accuracy of classication was assessed subsequently. Classication accuracies ranged from 68.7% to 75.5% (Table 3), which lies in the range of accuracy values of many maps obtained

by remote sensing and submitted to rigorous accuracy assessment (Zhu et al., 2000; Laba et al., 2002; Couturier et al., in press). According to the McNemars test (with p = 0.05), the values of accuracy of the three pixel-based classied images and objectbased classied images obtained with similarity of 24, 29, 34, 39, and 49 are not signicantly different. Therefore, the classied images with lower accuracy values were not considered further as the objective was to evaluate the variation of the landscape metrics derived from classied images with no signicant difference in accuracy. Compared with pixel-based, object-based approach did not always provide a greater accuracy of classication. This was particularly the case when applied to highly fragmented land cover types where it was impossible to ensure that every object contained only one class of pixels. This generalization may

Fig. 5. Classied image obtained by pixel-based maximum likelihood method followed by majority ltering, sieving (threshold of 5 pixels) and GIS operation to eliminate patches which size is below 22 pixels.

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224 Table 4 Statistics describing the variation of landscape metrics at landscape level. Structural feature Area/density/edge Metric NP LPI TE LSI AREA_MN AREA_AM GYRATE_MN GYRATE_AM SHAPE_MN SHAPE_AM FRAC_MN FRAC_AM PARA_MN PARA_AM CONTIG_MN CONTIG_AM PAFRAC TCA NDCA CORE_MN CORE_AM DCORE_MN DCORE_AM CAI_MN CAI_AM PROX_MN PROX_AM ENN_MN ENN_AM CONTAG PLADJ IJI DIVISION MESH SPLIT AI CONNECT COHESION PR SHDI SIDI MSIDI SHEI SIEI MSIEI Min. 5825 36.11 13716180 58.73 45.57 51593.70 153.49 8650.57 1.71 21.56 1.09 1.27 281.24 78.67 0.74 0.92 1.50 181540.71 4091.00 23.45 34580.67 36.76 13743.23 2.75 51.47 7835.73 6646.88 283.11 89.10 50.76 92.58 74.13 0.82 51483.11 5.59 92.70 4.38 99.75 8.00 1.65 0.74 1.35 0.79 0.85 0.65 Max. 7740 40.86 17321850 73.90 60.55 63268.16 167.11 9200.76 1.90 24.68 1.11 1.28 307.83 98.90 0.76 0.93 1.57 203025.15 5004.00 34.67 45334.30 49.37 41732.25 3.45 57.69 12392.99 15217.14 353.10 98.92 53.40 94.10 77.44 0.85 63136.14 6.85 94.22 4.72 99.78 8.00 1.69 0.76 1.41 0.81 0.86 0.68 Mean 6813.625 38.90 15484185 66.17 52.11 58070.77 159.15 8950.89 1.83 22.84 1.10 1.28 297.19 88.56 0.75 0.92 1.54 192526.11 4669.50 28.56 40545.64 41.48 26383.09 3.16 54.63 9536.06 11085.85 309.53 92.79 51.95 93.36 76.24 0.84 58023.31 6.11 93.48 4.51 99.76 8.00 1.67 0.75 1.38 0.81 0.86 0.67 S.D. 588.13 1.55 1331966.77 5.61 4.53 3979.88 4.50 216.68 0.09 1.11 0.01 0.00 10.30 7.48 0.01 0.01 0.03 8309.16 281.78 3.63 3841.88 4.25 10676.39 0.25 2.39 1264.53 3060.88 19.16 3.68 1.03 0.56 1.30 0.01 3959.77 0.42 0.56 0.14 0.01 0.00 0.02 0.01 0.03 0.01 0.01 0.01

1221

Coef. of variation (%) 8.63 3.97 8.60 8.47 8.69 6.85 2.83 2.42 4.84 4.85 0.74 0.36 3.47 8.44 1.20 0.72 1.79 4.32 6.03 12.69 9.48 10.25 40.47 7.88 4.37 13.26 27.61 6.19 3.96 1.97 0.60 1.71 1.34 6.82 6.86 0.60 3.13 0.01 0.00 1.23 0.92 1.98 1.23 0.92 1.98

Shape

Core area

Isolation/proximity

Contagion/interspersion

Connectivity

Diversity

cause mis-classication (Wang et al., 2004; Santos et al., 2007; Robertson and King, in press). These images (with no signicant difference in accuracy) are similar: Nearly 72% of the pixels exhibit the same class in the eight images, and 88% in six images or more. Fig. 5 shows the classied image obtained by pixel-based classication, followed by majority ltering and sieving with a threshold of 5 pixels. 5.2. Metric calculation and measurement error Tables 4 and 5 show the variation of metrics at the landscape and class level respectively. For area/density/edge metrics, and at the landscape level, the more stable metrics (variation coefcient <4%) were the largest patch index (LPI) and the mean radius of gyration (Gyrate_mn and Gyrate_am). The largest patch is a temperate forest area that is relatively compact and forms a major component of the landscape matrix in the north east of the image (Fig. 5). This patch does not suffer any important change in the different classied images. At the class level, LPI presents a very different behaviour depending on the land cover category. It is stable for more compact categories (forests, grassland) and

presents important variations for more fragmented categories (in particular orchards). Gyrate (mean distance between each cell in the patch and the patch centroid) seems to be a quite stable metric, except when weighted by patch size (gyrate_am) for categories which show important variations of patch areas among classied images. Among shape metrics, fractal dimension (FRAC) and perimeter area fractal dimension (PAFRAC) show less variation because they use the logarithm of perimeter and area, which is different from the shape index (SHAPE) and the perimeterarea ratio (PARA). The contiguity index (CONTIG, average contiguity value for the cells in a patch), is also an index with low variations. At the class level, the sensitivity of these indices depends on the compactness of each category. All core areas metrics exhibit important variations: the coefcients of variation range between 4% and 40% at the landscape level. At the class level, the coefcient of variation of the core area (CA) and of the disjoint core area (DCORE) reach 61% and 70% respectively for the orchard category. In order to estimate isolation/proximity, the Euclidean nearest neighbour distance (ENN) is less sensitive to the classication

1222

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

Table 5 Statistics describing the variation of landscape metrics at class level. Coefcient of variations describing the variation of landscape metrics at class level. The last four columns are respectively the minimum, maximum, mean and coefcient of variations of the values of the coefcient of variation for all the land cover categories. Metrics Cover/statistics Bare lands PLAND NP PD LPI TE ED LSI AREA_MN AREA_AM GYRATE_MN GYRATE_AM SHAPE_MN SHAPE_AM FRAC_MN FRAC_AM PARA_MN PARA_AM CONTIG_MN CONTIG_AM PAFRAC CPLAND NDCA CORE_MN CORE_AM DCORE_MN DCORE_AM CAI_MN CAI_AM PROX_MN PROX_AM ENN_MN ENN_AM CLUMPY PLADJ IJI DIVISION MESH SPLIT AI CONNECT COHESION 18.37 15.32 15.33 9.96 20.67 20.67 13.06 7.52 13.25 4.16 10.97 5.78 12.32 0.88 0.88 2.69 8.67 1.35 1.35 2.08 17.60 17.79 12.86 10.41 7.46 6.24 8.65 7.14 12.33 153.91 8.99 12.98 1.20 1.12 16.01 0.01 32.56 19.78 1.12 10.30 0.33 Grasslands 2.29 13.92 13.92 5.01 8.38 8.38 7.87 14.50 12.13 5.12 13.83 3.84 13.48 0.55 1.14 3.90 7.47 1.08 1.05 1.79 9.00 8.91 20.96 15.64 14.00 27.98 15.58 9.22 16.13 32.58 15.37 37.80 0.93 0.90 2.67 0.00 10.72 10.91 0.90 2.02 0.30 Human settlements 17.56 12.98 12.98 12.61 18.45 18.45 13.19 11.65 14.83 5.68 9.64 5.29 7.30 0.75 0.73 4.18 11.77 1.61 1.83 4.12 23.27 23.53 22.90 20.18 20.97 13.78 18.25 12.82 20.38 8.66 12.38 6.98 1.59 1.55 7.11 0.00 29.11 19.69 1.55 1.68 0.25 Irrigated agriculture 8.68 20.83 20.84 17.75 15.81 15.81 13.27 18.28 18.90 3.57 17.64 7.59 13.62 1.40 0.90 8.50 11.83 2.95 0.97 2.54 13.44 6.06 18.88 20.34 16.83 23.68 27.03 7.23 38.43 83.97 8.46 3.80 0.97 0.82 2.81 0.07 23.58 26.40 0.82 7.90 0.05 Rainfed agriculture 14.41 10.94 10.94 11.98 16.59 16.59 10.67 12.12 29.65 4.12 18.40 5.26 12.29 0.81 1.22 3.90 6.69 1.13 1.36 1.63 24.49 12.20 28.05 36.08 23.35 27.75 10.22 17.59 20.55 41.89 8.00 8.46 1.21 1.06 3.32 0.00 33.23 76.05 1.06 6.30 0.72 Orchards 11.05 8.18 8.19 35.20 5.99 5.99 4.43 18.84 56.28 3.39 34.12 3.28 19.01 0.48 1.04 3.47 6.90 1.39 0.67 0.49 20.62 6.15 29.71 61.49 22.53 70.54 9.31 8.68 33.34 34.62 16.16 3.36 0.64 0.65 21.53 0.05 75.17 33.55 0.64 3.70 0.21 Temperate forest 2.08 18.60 18.60 3.97 6.66 6.66 7.49 16.53 6.24 5.59 2.05 6.06 6.36 0.89 0.49 4.48 8.40 1.49 0.50 2.03 5.60 13.49 19.68 8.81 17.01 44.63 7.93 3.89 7.49 28.63 8.70 1.34 0.59 0.39 1.33 1.41 7.88 8.03 0.39 9.42 0.01 Tropical dry forest 2.51 9.43 9.44 3.79 8.30 8.30 8.34 12.16 10.13 5.46 6.22 6.73 9.94 0.98 0.72 4.59 8.64 1.55 0.57 1.41 4.41 4.15 13.86 9.87 7.40 7.72 7.33 2.85 20.17 31.06 9.59 1.87 0.60 0.50 3.32 0.11 12.55 11.02 0.50 1.94 0.03 Min. 2.08 8.18 8.19 3.79 5.99 5.99 4.43 7.52 6.24 3.39 2.05 3.28 6.36 0.48 0.49 2.69 6.69 1.08 0.50 0.49 4.41 4.15 12.86 8.81 7.40 6.24 7.33 2.85 7.49 8.66 8.00 1.34 0.59 0.39 1.33 0.00 7.88 8.03 0.39 1.68 0.01 Max. 18.37 20.83 20.84 35.20 20.67 20.67 13.27 18.84 56.28 5.68 34.12 7.59 19.01 1.40 1.22 8.50 11.83 2.95 1.83 4.12 24.49 23.53 29.71 61.49 23.35 70.54 27.03 17.59 38.43 153.91 16.16 37.80 1.59 1.55 21.53 1.41 75.17 76.05 1.55 10.30 0.72 Mean 9.62 13.78 13.78 12.53 12.61 12.61 9.79 13.95 20.18 4.64 14.11 5.48 11.79 0.84 0.89 4.46 8.80 1.57 1.04 2.01 14.81 11.53 20.86 22.85 16.19 27.79 13.04 8.68 21.10 51.91 10.96 9.57 0.97 0.87 7.26 0.21 28.10 25.68 0.87 5.41 0.24 Coef. of variation 66.43 29.75 29.74 77.47 43.53 43.53 31.27 25.70 74.98 18.82 64.82 24.20 31.71 31.34 25.29 36.47 21.26 35.16 41.29 48.44 49.83 53.78 26.93 73.45 36.12 71.88 49.20 51.26 45.53 83.55 27.84 117.70 34.22 40.02 95.79 220.50 71.58 80.46 40.15 61.08 91.44

approach than the proximity index (PROX) because it does not take into account the area of patches. At the class level, this last index has a coefcient of variation higher than 80% for bare lands and irrigated agriculture. The contagion/interspersion indices based on adjacencies, such as the percentage of like adjacencies (PLADJ), and the aggregation index (AI) are less sensitive than those indices derived from the area of the patches, such as the effective mesh size (MESH) and the splitting index (SPLIT). At class level, these two last indices have a large amount of variation particularly for the category orchards (with a coefcient of variation of 75.2 and 33.6% respectively). The connectivity index patch cohesion index (COHESION), based on area and perimeter of the patches, has less variation than the connectance index (CONNECT). CONNECT is derived from the number of functional joints between all the pairs of patches of the corresponding patch type within a distance specied by the user. The diversity index patch richness (PR) shows no variation because all the classied images have the same number of categories. Simpsons Diversity Index (SIDI) is less sensitive to change in scarce categories than the Shannons Diversity Index (SHDI) and therefore is more stable.

Fig. 6. Values of some metrics (as a proportion of maximum value) for the different similarity values (segmented images) and processing (pixel-based classication). NP number of patches, LPI largest patch index, LSI landscape shape index, Area_mn mean patch area, Gyrate_mn mean radius of gyration.

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224

1223

Generally, area-weighted means of metrics are not necessarily more stable than simple means of metrics, because the differences between the classied images obtained by the different approaches of classication and post-processing are not limited to small patches. In fact, the different approaches and parameters lead to the elaboration of images whose structure is different from one image to another. Small patches can appear or disappear, but also larger patches can be connected or disconnected. Therefore, both large and small patches can vary, depending on the classication and post-classication procedures applied. As expected, the segmentations carried out with larger similarity values produced images with smaller patches, and the landscape shape was less complex. However the effects of sieving and clumping are more difcult to predict (Fig. 6). Additionally, as pointed out by many authors (Riitters et al., 1995; Frohn, 1998; Gustafson, 1998) many of the indices used are correlated. For example, 77% of the metrics at landscape level have a coefcient of correlation above 0.5 (as absolute value). 6. Discussion and conclusion Many studies have aimed to quantify landscape pattern and/ or forest fragmentation over time (Fitzsimmons, 2003; Cayuela et al., 2006; Abdullah and Nakagoshi, 2007). For instance, Yang and Liu (2005) quantied landscape pattern and its change in an estuarine watershed in Florida and Alabama using two land use/ cover maps obtained through the classication of 1989 and 2002 Landsat images and the computing of landscape metrics. At the landscape level, they reported eight metrics (largest patch index, mean patch size, number of patches, area-weighted mean patch fractal dimension, core area coefcient of variation, total core area, modied Simpson diversity index and interspersion/ juxtaposition index) calculated for different spatial sub-units of the watershed. The coefcient of variation between 1989 and 2002 of the eight metrics ranged from 0.5% to 23.6%. The average value of the coefcient of variation of the largest patch index (calculated from the values obtained for the different sub-units), the index with less variation, was 3.9%. The index that exhibited the larger variation is the number of patch with an average coefcient of variation of 14.1%. At the class level, the variation of the value of the metrics between 1989 and 2002 depended upon the class and the spatial sub-units. The total core area of the category low density urban had an increase of 8666%, from 3 ha to 263 ha in one of the sub-units. However, for the other categories the majority of the metrics showed a variation of less than 20%. Therefore, the variations in the values of the indices due to the classication approach that we found in this work are of the same magnitude as the variations due to land cover/use change found by these authors. Consequently, the interpretation of the variation of landscape metrics of land use/cover maps of different dates obtained through the digital classication of images using (a) different approaches (pixel-based or object-based) or (b) different parameters when carrying out the segmentation of the image or the ltering of a pixel-based classied image must be taken with caution. This is because artifacts can produce signicant variations in the value of the metrics. In studies where metrics are compared between sub-regions extracted from the same image, and therefore classied with the same method/parameters, metrics which are more sensitive can be preferentially used. For instance, Kearns et al. (2005) removed the fractal dimension index from their analysis because it showed little variability, making it less useful for discriminating between the sub-watersheds they were comparing. For multidate comparison, when classication methods are distinct for the different dates, but other factors such as spatial resolution or anniversary date (and

thus vegetation phenology) are likely to play a role, more stable metrics should be preferred. However, metrics that are too stable, such as the patch richness, will not detect changes, and therefore a trade-off between robustness (stability) and sensitivity must be found. As shown in this study, variations in the value of landscape metrics can be the result of the approach used to analyze the images. Therefore, caution must be exercised in making meaningful comparisons and in detecting to what degree variations in metrics are really related to signicant changes in the landscape, and not to artefacts derived from methodological problems in their measurement. Comparison of landscape metrics must therefore be made with explicit knowledge of their sensitivity as demonstrated here. Acknowledgements This research has been supported by a grant from the SEMARNAT and CONACyT (project 2002-C01-0075, Fondo Sectori al de Investigacion Ambiental SEMARNAT-CONACyT). Special thanks go to project CONACYT-CONAFOR 2005-C02-14741 for supplying a PhD fellowship to the third author during writing-up of the paper. Final editing of the manuscript was done during a sabbatical stay of the rst author at the University of California n Santa Barbara with nancial support from the Direccio General de mico (DGAPA) and CONACyT. The Asuntos del Personal Acade authors also thank J. Liira for suggestions during the presentation of a preliminary version of this work at IUFRO Landscape Ecology International Conference held in Chengdu, China in September 2008, Dr M. McCall and three anonymous reviewers for their critical comments that greatly improved the earlier version of the manuscript. References
Abdullah, S.A., Nakagoshi, N., 2007. Forest fragmentation and its correlation to human land use change in the state of Selangor, peninsular Malaysia. Forest Ecology and Management 241, 3948. Baldwin, D.J.B., Weaver, K., Schnekenburger, F., Perera, A.H., 2004. Sensitivity of landscape pattern indices to input data characteristics on real landscapes: implications for their use in natural disturbance emulation. Landscape Ecology 19, 255271. Benson, B.J., MacKenzie, M.D., 1995. Effects of sensor spatial resolution on landscape structure parameters. Landscape Ecology 10, 113120. Bins, L.S., Fonseca, L.M.G., Erthal, G.J., Li, F.M., 1996. Satellite Image Segmentation: a region growing approach. In: INPE (Eds.), VIII Simposio Brasileiro de Sensor iamento Remoto, Salvador, Baha, Brazil, pp. 677680. Brown, D.G., Duh, J.-D., Drzyzga, S.A., 2000. Estimating error in an analysis of forest fragmentation change using North American Landscape Characterization (NALC) data. Remote Sensing of Environment 71, 106117. Camara, G., Souza, R.C.M., Freitas, U.M., Garrido, J., 1996. SPRING: Integrating remote sensing and GIS by object-oriented data modelling. Computers & Graphics 20, 395403. Cayuela, L., Benayas, J.M.R., Echeverra, C., 2006. Clearance and fragmentation of tropical montane forests in the Highlands of Chiapas, Mexico (19752000). Forest Ecology and Management 226, 208218. Couturier, S., Mas, J.F., Lopez, E., Bentez, J., Tapia, V., Vega, A., in press. Accuracy Assessment of the Mexican National Forest Inventory map: a study in four ecogeographical areas. Singapore Journal de Tropical Geography. Espindola, G.M., Camara, G., Reis, I.A., Bins, L.S., Monteiro, A.M., 2006. Parameter selection for region-growing image segmentation algorithms using spatial autocorrelation. International Journal of Remote Sensing 27, 30353040. Fitzsimmons, M., 2003. Effects of deforestation and reforestation on landscape spatial structure in boreal Saskatchewan, Canada. Forest Ecology and Management 174, 577592. Foody, G.M., 2004. Thematic map comparison: evaluating the statistical signicance of differences in classication accuracy. Photogrammetric Engineering & Remote Sensing 70, 627633. Forman, R.T.T., 1995. Land Mosaics: The Ecology of Landscapes and Regions. Cambridge University Press, Cambridge, 632 pp. Frohn, R.C., 1998. Remote Sensing for Landscape Ecology: New Metric Indicators for Monitoring, Modeling, and Assessment of Ecosystems. Lewis Publishers, Boca Raton, Florida, USA, 99 pp. Gustafson, E.J., 1998. Quantifying landscape spatial pattern: what is the state of the art? Ecosystems 1, 148156.

1224

J.-F. Mas et al. / Forest Ecology and Management 259 (2010) 12151224 Robertson, L.D., King, D.J., in press, Comparison of pixel- and object-based classication in land cover change mapping. International Journal of Remote Sensing. Santos, T., Tenedorio, J.A., Encarnac o, S., Rocha, J., 2007. Comparing pixel vs. object a based classiers for land cover mapping with Envisat-MERIS data. In: Bochenek, Z. (Ed.), 26th Annual Symposium of the European Association of Remote Sensing Laboratories (EARSeL). MillPress, Warsaw, Poland, p. 730. Saura, S., 2002. Effects of minimum mapping unit on land cover data spatial conguration and composition. International Journal of Remote Sensing 23, 48534880. Turner, M.G., ONeill, R.V., Gardner, R.H., Milne, B.T., 1989. Effects of changing spatial scale on the analysis of landscape pattern. Landscape Ecology 3, 153162. Vogt, P., Riitters, K.H., Estreguil, C., Kozak, J., Wade, T.G., Wickham, J.D., 2007. Mapping spatial patterns with morphological image processing. Landscape Ecology 22, 171177. Walter, V., 2004. Object-based classication of remote sensing data for change detection. Photogrammetry & Remote Sensing 58, 225238. Wang, L., Sousa, W.P., Gong, P., 2004. Intergration of object based and pixel based classication for mapping mangroves with IKONOS imagery. International Journal of Remote Sensing 25, 56555668. Yang, X., Liu, Z., 2005. Quantifying landscape pattern and its change in an estuarine watershed using satellite imagery and landscape metrics. International Journal of Remote Sensing 26, 52975323. Zhu, Z., Yang, L., Stehman, S.V., Czaplewski, R.L., 2000. Accuracy assessment for the U.S. Geological Survey regional land-cover mapping program: New York and New Jersey region. Photogrammetric Engineering and Remote Sensing 66, 14251435.

Haines-Young, R., Chopping, M., 1996. Quantifying landscape structure: a review of landscape indices and their application to forested landscapes. Progress in Physical Geography 20, 418445. Kearns, F.R., Kelly, N.M., Carter, J.L., Resh, V.H., 2005. A method for the use of landscape metrics in freshwater research and management. Landscape Ecology 20, 113125. Laba, M., Gregory, S.K., Braden, J., D, D.E., E, E.H., Fegraus, E., Fiore, J., DeGloria, S.D., 2002. Conventional and fuzzy accuracy assessment of the New York Gap Analysis Project land cover map. Remote Sensing of Environment 81, 443 455. McGarigal, K., Cushman, S.A., 2002. Comparative evaluation of experimental approaches to the study of habitat fragmentation effects. Ecological Applications 12, 335345. McGarigal, K., Cushman, S.A., Neel, M.C., Ene, E., 2002, FRAGSTATS: Spatial Pattern Analysis Program for Categorical Maps. University of Massachusetts, Amherst, computer software program available at www.umass.edu/landeco/research/ fragstats/fragstats.html. Meinel, G., Neubert, M., 2004. A Comparision of segmentation programs for high resolution remote sensing data. In: XXth ISPRS Congress, Istanbul, pp. 1097 1102. Qi, Y., Wu, J., 1996. Effects of changing spatial resolution on the results of landscape pattern analysis using spatial autocorrelation indices. Landscape Ecology 11, 3949. Riitters, K.H., ONeill, R.V., Hunsaker, C.T., Wickham, J.D., Yankee, D.H., Timmons, S.P., Jones, K.B., Jackson, B.L., 1995. A factor analysis of landscape pattern and structure metrics. Landscape Ecology 10, 2339.