You are on page 1of 9

ARTICLE IN PRESS

Computers & Geosciences 35 (2009) 18751883

Contents lists available at ScienceDirect

Computers & Geosciences


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

Optimization of scale and parametrization for terrain segmentation: An application to soil-landscape modeling
Lucian Dragut a,b,c,, Thomas Schauppenlehner d, Andreas Muhar d, Josef Strobl c,e, Thomas Blaschke c,f
a

Department of Geography and Geology, University of Salzburg, Hellbrunnerstrae 34, Salzburg 5020, Austria Department of Geography, West University of Timisoara, V. Parvan Blv. 4, Timisoara 300223, Romania c Z_GIS-Centre for Geoinformatics, University of Salzburg, Schillerstrae 30, Salzburg 5020, Austria d Institute of Landscape Development, Recreation and Conservation Planning, Department of Landscape, Spatial and Infrastructure Sciences, BOKU University of Natural Resources and Applied Life Sciences, Vienna, Peter Jordan-Str. 82, Vienna 1190, Austria e Austrian Academy of Sciences, Schillerstrae 30, Salzburg 5020, Austria f Research Studio iSpace, Schillerstrae 25, Salzburg 5020, Austria
b

a r t i c l e in fo
Article history: Received 25 April 2008 Received in revised form 8 September 2008 Accepted 7 October 2008 Keywords: Terrain segmentation Focal mean statistics Regression OBIA Curvature Soil productivity.

abstract
This paper presents a procedure to optimize parametrization and scale for terrain-based environmental modeling. The workow was exemplied on crop yield data, which is assumed to represent a proxy for soil productivity. Focal mean statistics were used to generate different scale levels of terrain derivatives by increasing the neighborhood size in calculation. The degree of association between each terrain derivative and crop yield values was established iteratively for all scale levels through correlation analysis. The rst peak of correlation indicated the scale level to be further retained. To select the best combination of terrain parameters that explains the variation of crop yield, we ran stepwise multiple regressions with appropriately scaled terrain parameters as independent variables. These techniques proved that the mean curvature, ltered over a neighborhood of 55 m, together with slope, made up the optimal combination to account for patterns of soil productivity. To illustrate the importance of scale, we compared the regression results of unltered and ltered mean curvature vs. crop yield. The comparison shows an improvement of R2 from a value of 0.01 when the curvature was not ltered, to 0.16 when the curvature was ltered within 55 55 m neighborhood size. The results were further used in an object-based image analysis environment to create terrain objects containing aggregated values of both terrain derivatives and crop yield. Hence, we introduce terrain segmentation as an alternative method for generating scale levels in terrain-based environmental modeling, besides existing per-cell methods. At the level of segments, R2 improved up to a value of 0.47. & 2009 Elsevier Ltd. All rights reserved.

1. Introduction Digital terrain analysis allows the generation of a suite of quantitative variables that reect geomorphic, climatic and hydrologic processes (McKenzie and Ryan, 1999). Therefore, gridded digital terrain models (DTMs) and their derivatives have been extensively used in terrain-based environmental modeling. More and more studies on the relationship between soil (re)distribution and topography utilize terrain attributes derived from digital elevation models. Recently, there is an increasing

Corresponding author at: Department of Geography and Geology, University of Salzburg, Hellbrunnerstrae 34, Salzburg 5020, Austria. Tel.: +43 662 8044 5293; fax: +43 662 8044 5260. E-mail address: lucian.dragut@sbg.ac.at (L. Dragut).

concern that conventional analyses using pixel-based approaches do not incorporate the concept of landscape features or real-world objects (Smith et al., 2007; Wulder et al., 2008). While standard applications consider terrain information as expressed through points, objects have been proposed as alternatives (Blaschke and Strobl, 2003; Deng, 2007). Such discrete units have been proposed particularly for soil-landscape modeling purposes. Conacher and Dalrymple (1977) introduced the nine-unit landsurface model, while Pennock and Corre (2001) summarized the concept and procedures of landform segmentation. Both approaches assume that functionally distinct objects, as resulted from land surface partition, are linked with soil formation processes. An advantage of landform objects over point representation is that the topological information can be added to models. Recently, object-based image analysis (OBIA) has proved a valuable framework to integrate the idea of representing land

0098-3004/$ - see front matter & 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.cageo.2008.10.008

ARTICLE IN PRESS
1876 L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883

surface through discrete entities. Until some years ago, OBIA techniques were mainly used to produce image objects, which were then either extracted or classied. Its application to terrain data has started being consolidated theoretically (Minar and Evans, 2008; Strobl, 2008), in parallel with developments of landform classication systems (Dragut and Blaschke, 2006, 2008; Stepinski et al., 2007). Besides general potential benets of the object-based approach (for an overview see Blaschke et al., 2008) compared to per-pixel techniques (Blaschke and Strobl, 2001; Wulder et al., 2008), OBIA has the potential of improving terrain representation by considering anisotropy. Schmidt and Andrew (2005) have identied the isotropic characterization of surface in terrain analysis as inadequate, since the real land surfaces show spatial anisotropy. This problem, generated by xed shape and uniform size of image objects, can be tackled through segmentation techniques (Addink et al., 2007). So far, both per-pixel and object-based classications of discrete objects have been proposed as generic systems, except for classications rooted in the parametric approach of Speight (1974). To turn them into specic classication system for terrain-based environmental modeling, parametrization plays an important role, besides scale. The issue of parametrization is particularly a challenge in OBIA, since the number of possible tessellations has increased innitely (Addink et al., 2007) through the process of image segmentation. Since the terrain objects produced through segmentation are strongly dependent on the number of and information contained by layers (e.g., terrain parameters), the optimization of data input for OBIA deserves an appropriate consideration. Either pixel or object-based, terrain-based environmental modeling requires nding the terrain parameters that explain the variable of interest. This is typically performed with various statistical techniques, usually pixel-based and for readily available DTM resolutions. Thus, the analysis is tied to the DTM resolution, which induces the scale of modeling. However, there is an increasing amount of literature underpinning the impact of the mismatch of scales between the target variable and the explanatory variables. Both elevation and its derivatives have scale-related environmental roles; therefore scale should become part of terrain attributes (Deng, 2007). As demonstrated in a number of studies (Speight, 1977; Bian and Walsh, 1993; Florinsky and Kuryakova, 2000; Schmidt and Andrew, 2005), the relationships between terrain parameters and physical processes are meaningful at specic scales. Technically, this means that the standard procedure used to produce terrain parameters, e.g., applying a 3 3 moving window on a DTM at a given resolution might not be appropriate when terrain variables are related to landscape properties. It is usually not known what the optimal scale is. If the spatial resolution of the DEM and the respective parameters derived from it are relatively coarse, then usually the nest possible spatial resolution is used for the following analyses. If the available resolution is very ne, methods are needed to hit the optimal scale within the analysis process. As scale is a window of perception (Marceau, 1999) there is no right or wrong scale per se. The appropriate scale for the analyses has to be found out relative to the data sets and the purpose of the study. To nd the right scale in relational studies, statistical analyses have been proposed as means of exploring relationships between target variables (soil or vegetation) and terrain variables at various scales. These scales have been simulated by (1) constantly increasing either the cell size of DTMs (Bian and Walsh, 1993; Florinsky and Kuryakova, 2000; Arrell et al., 2007) or the neighborhood size (Fisher et al., 2004; Schmidt and Andrew, 2005) in computation of terrain parameters using the Woods (1996) method; or (2) ltering variables (Moran and Bui, 2002; Deng et al., 2007; Jordan, 2007; Grinand et al.,

2008). These approaches are theoretically framed within the concept of characteristic scales associated with scale breaks (Bian and Walsh, 1993; Wu and Li, 2006). In this study we propose a hybrid approach for parametrization and scaling of terrain data, using both per-pixel and objectbased techniques. The overall goal of this paper is to develop a procedure for optimizing the production of input data for terrain segmentation. The procedure shall enhance terrain-based environmental modeling through selection of the most suitable terrain derivatives, at the appropriate scale for the process considered. The workow is exemplied through the soil-landscape modeling.

2. Material and methods The study area is located about 7 km east of Vienna in the village of Rutzendorf (municipality Gro-Enzersdorf, Lower Austria) in the Marchfeld area, an alluvial plain of the Danube River (Fig. 1). The area is under intense agricultural use with high fertile chernozems as dominant soil group. Although very at, with only 2 m elevation difference within the 20.74 ha study area, and homogeneous in terms of parent material, the area shows large differences in soil quality, ranging from 49 to 87 index points on a scale from 0 to 100 in the Austrian soil quality index for cropping (Ackerzahl). This study was conducted using data from a research project on long-term effects of the conversion of an experimental farm in Eastern Austria to organic farming (Freyer et al., 2007). One particular task within this project was to provide detailed information on the spatial variability of soil quality for the selection of locations for long-term monitoring plots. A DTM at 1 1 m resolution was derived from stereo aerial photos at a vertical accuracy of less than 20 cm. The DTM was resampled to 5 m spatial resolution using bilinear interpolation to reduce noise (Evans, 1972; Guth, 1995; Jordan, 2003). The resulting DTM consisted of 120 columns by 133 rows, with elevation values ranging between 152 and 154 m (Fig. 1B). Crop yield data was available for the study area (Fig. 1C). The data were collected with an AFS Yield Monitor from CASE IH Agriculture. The 5 m average sample distance of the yield data results from the harvester width and the time between two measurements. The used GPS-Receiver was a DGPS-device from Trimble, with a position lapse between 0.51 m. A ow measuring device was used to quantify the amount of grain in the hopper. Given that the device is calibrated correctly, an accuracy of 73% is attainable with this setup.1 The raw data from the harvester contained many void values from different sources which could be identied as calibrating error, unknown cutting width, time lapse, position lapse, defects in the crop transportation system and measuring errors (Blackmore and Moore, 1999). These voids were removed using a distance function between two points in a row to eliminate points which were too close together (Robinson and Metternicht, 2005). Because of the unsystematic structure of the errors in the turnaround area, these points were not used for validation. Eventually, more than 10.500 points were retained for the analysis (Schauppenlehner, 2008), with harvest values ranging between 1680 and 6190 kg/ha. Data available for soil taxation purposes was used for evaluating the outputs. In Austria, a systematic soil quality survey with the goal to guarantee a uniform base for taxation purposes is carried out by the scal authorities; these data offer detailed soil information on all agricultural areas in Austria at high resolution (50 m average sample distance). The data acquisition process
1 Prairie Agricultural Machinery Institute, 1999. http://pami.ca/pdfs/reports_ reasearch_updates/745_a_comparison_of_three_popular_yield_monitors_and_gps_receivers.pdf

ARTICLE IN PRESS
L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883 1877

Fig. 1. Location of study area in Austria (A). Digital terrain model (B). Crop yield point data, with values representing harvest in kg/ha (C). Classied eld map for taxation purposes, with values representing soil quality rating (D).

consists in collecting soil samples at the dened raster, rating them on a scale from 0 to 100, and extracting areas of homogenous soil quality classes based on expert knowledge and relief through manual delineation (Wagner, 2001). The manual delineation relies on a strict estimation framework, where a soil quality range can be identied by using parameter values for soil texture, soil condition and soil genesis (Wagner, 2001). For our study, both the point data (79 samples) and the classied eld map were available in the form of hand-drawn paper maps and structured eld protocols, which needed to be digitized and georeferenced. The procedure for optimizing the production of input data for terrain segmentation is represented in Fig. 2. In this particular study, the optimization is relative to relationships between terrain parameters and soil productivity. Yield data was considered the best approximation of the soil productivity, yet assuming its limitations (Robinson and Metternicht, 2005; Vitharana et al., 2008). Using this kind of data gives the advantage of avoiding reduced sample size, which is a constant problem in pedology (Gessler et al., 2000), yet the statistical signicance increases. The statistical relationships between yield data and terrain derivatives were used to select (1) the scale at which each terrain parameter correlates the best with crop yield data; and (2) the specic

combination of terrain parameters that predicts the best soil productivity patterns. Based on the results of the statistical analyses, terrain segmentation techniques were applied. Objects resulting from terrain segmentation were compared with manual delineation produced for soil taxation purposes. The main steps in developing the workow were as follows.

2.1. Production of terrain derivatives Digital models of elevation, slope, Terrain Wetness Index (TWI), prole, plan and absolute curvatures were derived from the DTM. TWI is dened as ratio between catchment area and slope and was calculated with integrated land and water information system (ILWIS), developed by International Institute for Geo-Information Science and Earth Observation (ITC, The Netherlands) using a multiple ow directions algorithm (Quinn et al., 1991). For this area, TWI values range from 4.8 to 15.1. All other parameters were derived using ArcGIS in a standard 3 3 neighborhood size. Absolute curvature, plan curvature and prole curvature values range between 6.5 to 4.4, 4.5 to 1.8, and 5.2 to 3.2, respectively. Here, we use the term absolute curvature to differentiate from plan and prole curvatures. We did not study

ARTICLE IN PRESS
1878 L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883

Fig. 2. Flow diagram of methods used to produce object primitives.

the inuence of the slope aspect on soil productivity, since the study area is very at (slope values range between 0 and 5.1 degrees). 2.2. Simulation of scales for curvatures In the rst part of this work, we explored the potential of neighborhood statistics to simulate various scales of curvatures in a similar manner as proposed by Deng et al. (2007). The focal mean statistics on increasing neighborhood window was calculated for all curvatures. Thus, a new value is assigned for each cell as the mean of its n neighbors, where n is controlled by the size of the squared window (e.g., 9 neighbors for a 3 3 window, 25 for a 5 5 window and so on). Scales were simulated by increasing the window size with odd numbers, starting with 3. At each scale, correlation analysis between the respective curvature and crop yield values were performed. This recursive process stopped when the rst peak of correlation was reached. This procedure has not been applied for elevation and slope due to their extremely low range of values. In addition, we did not apply focal mean statistics on TWI, since the values of this compound index make sense only for the cells for which it was calculated. The lack of sensitivity of this index to scale transform analysis has been demonstrated elsewhere (Yong et al., 2008). This procedure allows the identication of the most appropriate scale at which curvatures (plan, prole and absolute) t crop yield data. 2.3. Parametrization To identify which combination of terrain parameters has the highest potential of explaining the pattern of soil productivity, we used stepwise multiple regression analysis. All curvatures at the scales of correlation peaks, together with elevation, slope and TWI were used as independent variables to predict the crop yield. The combination of terrain parameters as indicated in the model maximize the R2. 2.4. Terrain segmentation We imported the digital models of terrain parameters most explanatory of soil fertility patterns (here absolute curvature and slopes), plus elevation into Deniens Developer 7 to perform the OBIA. For practical purposes, curvature values were multiplied by 100 as Zevenbergen and Thorne (1987) recommended. Image

Table 1 Pearsons correlation of terrain parameters with crop yield data.


CY Elevation Slope CY 1 0.192 TWI Prole Plan Curvature Prole11 Plan11 K11 0.341 0.219 0.396

0.156 0.312 0.090 0.082 0.098

All correlations were signicant at 0.01 level.

segmentation groups cells into objects by pair-wise clustering, beginning with single-pixel objects (Baatz and Schape, 2000). The regionalization process stops when the heterogeneity threshold (a dimensionless value called scale parameter) set by the user is exceeded. The composition of heterogeneity can be adjusted by allocating a desired percentage of shape information besides spectral information (here the term spectral information refers to cell values). There is no method for objectively selecting the best combination of parameters for segmentation; the user must search for the most appropriate combination in a recursive manner, by trial and error (Walker and Blaschke, 2008). The segmented objects (object primitives, as they have not assigned any meaning yet) were created based exclusively on absolute curvature, slope and elevation layer values (without shape parameters), with equal weights, using a scale parameter of 5. Although the elevation was not a good predictor of crop yield, it was used in the segmentation process merely to optimize the shape of object primitives according to the terrain pattern. The segmentation produced 36 object primitives for the study area.

3. Results and discussion Terrain parameters with curvatures both ltered and unltered were correlated with crop yield data (Table 1). CY represents crop yield data. Prole11, Plan11, and K11 represent the values of prole, plan and absolute curvatures calculated with focal mean function over of a neighborhood of 11 11 cells (equal to 55 55 m). Table 1 shows that curvatures derived in a standard procedure have the worst correlation with crop yield data amongst all terrain derivatives. However, when scaled as described above, their correlations increased considerably, so that prole and absolute curvatures (Prole11 and K11 in Table 1) outperformed the TWI. As marked in Table 1, absolute curvature at the scale 11 correlates best with crop yield data. Fig. 3 portrays the correlation variation of absolute curvature across scales. It is visible that the correlation increases up to the scale level created over a neighborhood of 11 11 cells, then starts decreasing.

ARTICLE IN PRESS
L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883 1879

Fig. 3. Pearsons correlation between crop yield data and absolute curvature at scales between 1 (without any focal mean statistics, indicated by K) and 13 (K13).

Fig. 4. Scatter plot of crop yield values vs. absolute curvature: non-smoothed (left) and smoothed (right). Crop yield values are expressed in dt/ha.

Therefore, scale level K11 (see Fig. 3) was retained for further analysis. The same procedure was applied for plan and prole curvatures, which reached the maximum correlation at the same scale level as the absolute curvature (11 11 cells neighborhood). The regression between crop yield data and the absolute curvature increased 16 times, from an R2 value of 0.01 when the curvature was not ltered, to 0.16 when the curvature was ltered with 55 55 m neighborhood size (Fig. 4). The increase in both Pearsons value and R2 with the neighborhood size of ltering reveals the effect of spatial scale upon relationships between the soil productivity patterns and terrain derivatives. This is because geomorphic phenomena are scale-dependent, commonly irregular even random at high resolution, but exhibiting order at coarser scales (Iwahashi et al.,

2001, pp. 853). Fig. 5 illustrates the above statement through comparison between maps of absolute curvature not ltered (left) and ltered (right). The local variation displayed by the original map of curvature was smoothed at coarser scale, where pattern of surface curvature explains better the soil productivity. The physical meaning of scaling terrain derivatives appears thus clear: topography inuences soil processes only within specic ranges of landforms. Similar scale behavior of terrain parameters has been found in relationship to soil moisture or vegetation properties (Bian and Walsh, 1993; Florinsky and Kuryakova, 2000; Deng et al., 2007). Here, surface curvature of approximately 55 m holds the highest inuence on soil productivity, through moisture availability. If not errors, more localized curvatures are likely obliterated by other processes inuencing soil productivity. The

ARTICLE IN PRESS
1880 L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883

Fig. 5. Maps of absolute curvature non-smoothed (left) and smoothed with focal mean statistics in a window of 11 cells (right).

Table 2 Model summary of stepwise regression analysis with terrain parameters as independent variables and crop yield as dependent variable. Model R R2 Adjusted R2 Standard error of the estimate 3.98970283 3.91657521 3.90915375 3.90803016

1 2 3 4
a

0.396a 0.433b 0.436c 0.437d

0.157 0.187 0.191 0.191

0.157 0.187 0.190 0.191

Predictors: (Constant), Curvature 11. Predictors: (Constant), Curvature 11, Slope. Predictors: (Constant), Curvature 11, Slope, Prole curvature 11. d Predictors: (Constant), Curvature 11, Slope, Prole curvature 11, TWI.
b c

same parameter was found a suitable indicator of variations in topography affecting crop yield by Timlin et al. (1998). Shary et al. (2002) explained theoretically the impact of curvature on memory in soils. Florinsky and Kuryakova (2000) showed that the curvature was the most suitable parameter to explain the topographic control on soil moisture. For the climatic conditions of this area, soil productivity depends mainly on moisture conditions in the soil. Gessler et al. (2000) also found strong relationships between several soil properties (soil depth, A horizon depth, Cmass, NPP) and terrain concavity, which was evaluated using the compound topographic index (CTI). To estimate the most appropriate combination of terrain parameters likely to inuence soil productivity in this area, we ran stepwise multiple regression analysis results (Table 2). All terrain parameters, with curvatures at peaks of correlation, were used as independent variables to predict the crop yield. In Table 2, the second model with absolute curvature and slope as predictors shows the most signicant improvement from adjusted R2 of 0.157 to 0.187. Models where more predictors were added only improved the regression with less than 0.4%. Based on results of regression analysis, we used absolute curvature at the scale K11 and slope as layers in terrain

segmentation. As mentioned in the previous chapter, we added the elevation layer as well. Results of terrain segmentation are represented in Fig. 6 (left). Object primitives have attached mean values of terrain derivatives that were used in segmentation. Here, the mean values of absolute curvature are symbolized on the map. We compared segmentation results with manually delineated polygons produced for taxation purposes (Fig. 6 right). Compared to manual delineation, terrain classication generated more complex partition of the surface (Fig. 6), with higher number of polygons (36 compared to 26). To assess the results of terrain segmentation, we plotted mean crop yield values against mean values of prole curvature and compared the graphic with the scatter plot of mean crop yield values vs. soil productivity values assigned to the manually delineated polygons included in the original eld assessment map of the scal authority (Fig. 7). In both cases crop yield values were aggregated within the polygons in Fig. 6. The need for aggregating both dependent and independent variables in relational studies was pointed out by Bian and Walsh (1993). In the cited work, vegetation index and the explanatory topographic variables were aggregated in a pixel-based framework to reach the characteristic scale where the two sets of variables were most highly related (Bian and Walsh, 1993, pp. 10). Thus, local variation was smoothed into a categorical representation (Moran and Bui, 2002). Here, we introduced an alternative procedure in an OBIA environment. When calculated for segmented objects, R2 of relationship between curvature and crop yield reached a value of 0.47 (Fig. 7 left). Therefore, high crops are signicantly associated to segments representing surface concavities, while surface convexities predict patterns of lower crops (see also Fig. 6). Unexpectedly, relationship between values of polygons produced for taxation purposes and mean values of crop yield attached to these polygons is characterized by R2 of 0.19 (Fig. 7 right). Moreover, the sense of relationship is negative. This suggests that soil taxation data failed in predicting the crop yield in this area. However, when plotting the soil sample values (Ackerzahl) in a one-to-one association, the results are better,

ARTICLE IN PRESS
L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883 1881

Fig. 6. Comparison between object primitives (left) and manually delineated polygons (right). Polygon values represent: mean absolute curvature (left) and soil quality rating based on expert knowledge (right).

Fig. 7. Scatter plot of mean crop yield values vs. mean values of absolute curvature (left) and soil quality rating based on expert knowledge (right). Crop yield values are expressed in dt/ha.

with a result of R2 of 0.34 (Fig. 8). For all 79 soil samples within study area, the closest crop yield values were attached and then plotted against each other. The difference in predicting crop yield is likely due to human errors in drawing polygons (see Speight (1977), p. 162 for criticism on manually extrapolating point data to mapped areas). For instance, although including one of the most productive units, the polygon with value 86 in the central area of Fig. 6 crosses several terrains with lower crop yield values. Such errors are particularly expected when visually transforming smooth terrain transitions into categorical units. In contrast, terrain transitions expressed as values could be better handled

using computer techniques. Thus, we argue that the tessellation generated through terrain segmentation approach can be successfully integrated into expert knowledge procedures that aim at delineating soil quality units with the help of soil samples. Object primitives would enhance photo-interpretation in soil survey (Hengl and Rossiter, 2003) or could be used as basis for sampling (Minasny et al., 2007) or for interpolation of samples.2

2 Schauppenlehner et al. (2008). http://www.cosis.net/abstracts/EGU2008/ 05303/EGU2008-A-05303.pdf?PHPSESSID=41c8cb9b6ed0f07b8fd59b71c56dc6f6.

ARTICLE IN PRESS
1882 L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883

As found in some previous papers (Timlin et al., 1998; Florinsky and Kuryakova, 2000; Gessler et al., 2000), surface curvature (appropriately scaled!) was the best predictor of soil productivity through the control it holds on soil moisture. This study also conrmed previous research (Bian and Walsh, 1993; Florinsky and Kuryakova, 2000; Moran and Bui, 2002; Fisher et al., 2004; Schmidt and Andrew, 2005; Deng, 2007; Deng et al., 2007; Grinand et al., 2008) on the impact of scale on terrainlandscape modeling. The analysis undertaken shows an improvement of relationships between crop yield data and the curvature from an R2 value of 0.01 when the curvature was not ltered to 0.16 and when the curvature was ltered with 55 55 m neighborhood size. This value improved even more, up to 0.47, when calculated for segmented objects. In the last case, both crop yield data and total curvature display mean values as resulted from clustering cells into object primitives. Hence, we introduce terrain segmentation as an alternative method for generating scale levels in terrain-based environmental modeling, besides existing per-cell methods.

Fig. 8. Scatter plot of crop yield values vs. soil samples quality index (Ackerzahl). Crop yield values are expressed in dt/ha.

Acknowledgements The data used for this study were taken from the research project MUBIL (monitoring the conversion to organic farming), funded by the Austrian Federal Ministry of Agriculture, Forestry, Environment and Water Management; project leader: Prof. Bernhard Freyer, Institute of Organic Farming, BOKU University of Natural Resources and Applied Life Sciences Vienna. The work of Lucian Dragut was supported by a Marie Curie Fellowship of the European Commission (CLUE, MEIF-CT-2005-009532) and a Stand-alone project funded by the Austrian Science Fund (SCALA, FWF-P20777-N15). We are grateful to two anonymous reviewers for some very helpful comments. References
Addink, E.A., de Jong, S.M., Pebesma, E.J., 2007. The importance of scale in objectbased mapping of vegetation parameters with hyperspectral imagery. Photogrammetric Engineering & Remote Sensing 73, 905912. Arrell, K.E., Fisher, P.F., Tate, N.J., Bastin, L., 2007. A fuzzy c-means classication of elevation derivatives to extract the morphometric classication of landforms in Snowdonia, Wales. Computers & Geosciences 33, 13661381. Baatz, M., Schape, A., 2000. Multi-resolution segmentation. An optimization approach for high-quality multi-scale image segmentation. In: Strobl, J., Blaschke, T., Griesebner, G. (Eds.), Angewandte Geographische Informationsverarbeitung XII. Wichmann Verlag, Heidelberg, pp. 1223. Blaschke, T., Lang, S., Hay, G.J. (Eds.), 2008. Object-Based Image Analysis. Spatial Concepts for Knowledge-Driven Remote Sensing Applications. Springer, Heidelberg, Berlin, p. 818. Bian, L., Walsh, S.J., 1993. Scale dependencies of vegetation and topography in a mountainous environment of Montana. Professional Geographer 45, 111. Blackmore, S., Moore, M., 1999. Remedial correction of yield map data. Precision Agriculture 1, 5366. Blaschke, T., Strobl, J., 2001. Whats wrong with pixels? Some recent developments interfacing remote sensing and GIS. GISZeitschrift fur Geoinformationssysteme 14, 1217. Blaschke, T., Strobl, J., 2003. Dening landscape units through integrated morphometric characteristics. In: Buhmann, E., Ervin, S. (Eds.), Landscape Modeling: Digital Techniques for Landscape Architecture. Wichmann Verlag, Heidelberg, pp. 104113. Conacher, A.J., Dalrymple, J.B., 1977. (Eds.), The Nine-Unit Landsurface Model: An Approach to Pedogeomorphic Research (special issue). Geoderma 18 (12). Deng, Y., 2007. New trends in digital terrain analysis: landform denition, representation, and classication. Progress in Physical Geography 31, 405419. Deng, Y., Chen, X., Chuvieco, E., Warner, T., Wilson, J.P., 2007. Multi-scale linkages between topographic attributes and vegetation indices in a mountainous landscape. Remote Sensing of Environment 111, 122134. Dragut, L., Blaschke, T., 2006. Automated classication of landform elements using object-based image analysis. Geomorphology 81, 330344. Dragut, L., Blaschke, T., 2008. Terrain segmentation and classication using SRTM data. In: Zhou, Q., Lees, B., Tang, G.A. (Eds.), Advances in Digital Terrain Analysis, Series Lecture Notes in Geoinformation and Cartography. Springer, New York, pp. 141158.

The methodology introduced in this paper was applied to soil data. Nevertheless, it can be used in the analysis of topographic inuence on other environmental variables (e.g., vegetation) as a procedure to optimize terrain parametrization in connection with the characteristic scale at which relationships between variables (dependent and independent) are best possible. The combination of terrain parameters with the greatest impact on the target environmental variable, under the particular conditions of the study area and available datasets, is identied using stepwise regression. The optimal scale indicates the particular level at which terrain parameters really inuence the environmental process of interest. Thus, the techniques reveal the most basic structures of both terrain and environmental data, by eliminating more localized variations, which are likely to be driven by other factors. Before integrating such an approach within a routine procedure, a thorough comparative assessment of the scale/multi-scale methods (i.e. increasing grid size, increasing neighborhood size, ltering or OBIA) is required. This is the subject of our future research.

4. Conclusions The study developed a hybrid procedure to optimize parametrization and scale for terrain-based environmental modeling. Per-cell approach was carried out to select the optimal combination of terrain parameters, at appropriate scale for the environmental variable of interest, which were then used as input for OBIA. The workow was exemplied on crop yield data, which has been used as reference dataset, with the purpose of producing an appropriate tessellation for predicting patterns of soil productivity. The methodology introduced in this study has been proven to be effective in delineating terrain objects that predict better patterns of soil productivity compared to standard methods used by the Austrian scal authorities for taxation purposes. It has been proved that research on the relationship between soil productivity and topography can be improved by the ease with which terrain attributes can be derived from digital elevation models when appropriate scale parameters are used.

ARTICLE IN PRESS
L. Dragut et al. / Computers & Geosciences 35 (2009) 18751883 1883

Evans, I.S., 1972. General geomorphometry, derivatives of altitude, and descriptive statistics. In: Chorley, R.J. (Ed.), Spatial Analysis in Geomorphology. Methuen, London, pp. 1791. Fisher, P., Wood, J., Cheng, T., 2004. Where is Helvellyn? Fuzziness of multi-scale landscape morphometry. Transactions Institute of British Geographers 29, 106128. Florinsky, I.V., Kuryakova, G.A., 2000. Determination of grid size for digital terrain modelling in landscape investigations exemplied by soil moisture distribution at a micro-scale. International Journal of Geographical Information Science 14, 815832. Freyer, B., Surbock, A., Friedel, J.K., Heinzinger, M., 2007. Zukunft einer okologisch genutzten Agrarlandschaft (The future of an ecologically used agricultural landscape). GAIA. Ecological Perspectives for Science and Society 16, 158160. Gessler, P.E., Chadwick, O.A., Chamran, F., Althouse, L., Holmes, K., 2000. Modeling soil-landscape and ecosystem properties using terrain attributes. Soil Science Society of America Journal 64, 20462056. Grinand, C., Arrouays, D., Laroche, B., Martin, M.P., 2008. Extrapolating regional soil landscapes from an existing soil map:sampling intensity, validation procedures, and integration of spatial context. Geoderma 143, 180190. Guth, P.L., 1995. Slope and aspect calculations on gridded digital elevation models: examples from a geomorphometric toolbox for personal computers. Zeitschrift fur Geomorphologie Supplementband 101, 3152. Hengl, T., Rossiter, D.G., 2003. Supervised landform classication to enhance and replace photo-interpretation in semi-detailed soil survey. Soil Science Society of America Journal 67, 18101822. Iwahashi, J., Watanabe, S., Furuya, T., 2001. Landform analysis of slope movements using DEM in Higashikubiki area, Japan. Computers & Geosciences 27, 851865. Jordan, G., 2003. Morphometric analysis and tectonic interpretation of digital terrain data: a case study. Earth Surface Processes and Landforms 28, 807822. Jordan, G., 2007. Adaptive smoothing of valleys in DEMs using TIN interpolation from ridge elevations: an application to morphotectonic aspect analysis. Computers & Geosciences 33, 573585. Marceau, D., 1999. The scale issue in the social and natural sciences. Canadian Journal of Remote Sensing 25, 347356. McKenzie, N.J., Ryan, P.J., 1999. Spatial prediction of soil properties using environmental correlation. Geoderma 89, 6794. Minar, J., Evans, I.S., 2008. Elementary forms for land surface segmentation: the theoretical basis of terrain analysis and geomorphological mapping. Geomorphology 95, 236259. Minasny, B., McBratney, A.B., Walvoort, D.J.J., 2007. The variance quadtree algorithm: use for spatial sampling design. Computers & Geosciences 33, 383392. Moran, C.J., Bui, E.N., 2002. Spatial data mining for enhanced soil map modelling. International Journal of Geographical Information Science 16, 533549. Pennock, D.J., Corre, M.D., 2001. Development and application of landform segmentation procedures. Soil and Tillage Research 58, 151162. Quinn, P., Beven, K., Chevallier, P., Planchon, O., 1991. The prediction of hillslope ow paths for distributed hydrological modelling using digital terrain models. Hydrological Processes 5, 5979. Robinson, T.P., Metternicht, G., 2005. Comparing the performance of techniques to improve the quality of yield maps. Agricultural Systems 85, 1941. Schauppenlehner, T., 2008. Geostatistische Analysen zur Integration von Gelande modellen und Bodenschatzungsdaten fur verbesserte digitale Bodenkarten am

Beispiel einer alluvialen Landschaft (Geostatistical analysis to integrate digital elevation models and soil taxation data for the development of improved digital soil maps in an alluvial ood plain). Unpublished PhD Dissertation, BOKU University of Natural Resources and Applied Life Sciences, Vienna, pp. 133. Schmidt, J., Andrew, R., 2005. Multi-scale landform characterization. Area 37, 341350. Shary, P.A., Sharaya, L.S., Mitusov, A.V., 2002. Fundamental quantitative methods of land surface analysis. Geoderma 107, 132. Smith, G., Beare, M., Boyd, M., Downs, T., Gregory, M., Morton, D., Brown, N., Thomson, A., 2007. UK land cover map production through the generalisation of OS mastermaps. The Cartographic Journal 44, 276283. Speight, J.G., 1974. A parametric approach to landform regions. In: Brown, E.H., Waters, R.S. (Eds.), Papers in Honour of D.L. Linton, Progress in Geomorphology, Institute of British Geographers Special Publication no. 7, London, pp. 213230. Speight, J.G., 1977. Landform pattern description from aerial photographs. Photogrammetria 32, 161182. Stepinski, T.F., Ghosh, S., Vilalta, R., 2007. Machine learning for automatic mapping of planetary surfaces. In: Proceedings 19th Innovative Applications of Articial Intelligence Conference, San Jose, California, USA, pp. 718. Strobl, J., 2008. Segmentation-based terrain classication. In: Zhou, Q., Lees, B., Tang, G.A. (Eds.), Advances in Digital Terrain Analysis, Series Lecture Notes in Geoinformation and Cartography. Springer, New York, pp. 125139. Timlin, D.J., Pachepsky, Y., Snyder, V.A., Bryant, R.B., 1998. Spatial and temporal variability of corn grain yield on a hillslope. Soil Science Society of America Journal 62, 764773. Vitharana, U.W.A., Van Meirvenne, M., Simpson, D., Cockx, L., De Baerdemaeker, J., 2008. Key soil and topographic properties to delineate potential management classes for precision agriculture in the European loess area. Geoderma 143, 206215. Wagner, J., 2001. Bodenschatzung in Osterreich (Soil taxation in Austria). Bodenaufnahmesysteme in Osterreich: Mitteilungen der Osterreichischen Bodenkundlichen Gesellschaft 62, 69104. Walker, J., Blaschke, T., 2008. Object-based landcover classication for the Phoenix metropolitan area: optimization vs. transportability. International Journal of Remote Sensing 29, 20212040. Wood, J., 1996. The geomorphological characterisation of digital elevation models, PhD Dissertation, University of Leicester, England, /http://www.soi.city.ac.uk/ $jwo/phd/S, [accessed 23 March 2009], 193p. Wu, J., Li, H., 2006. Concept of scale and scaling. In: Wu, J., Jones, K.B., Li, H., Loucks, O.L. (Eds.), Scaling and Uncertainty Analysis in Ecology: Methods and Applications. Springer, Dordrecht, pp. 315. Wulder, M.A., White, J.C., Hay, G.J., Castilla, G., 2008. Pixels to objects to information: Spatial context to aid in forest characterization with remote sensing. In: Blaschke, T., Lang, S., Hay, G.J. (Eds.), Object-Based Image Analysis. Spatial Concepts for Knowledge-Driven Remote Sensing Applications. Springer, Heidelberg, Berlin, pp. 345363. Yong, B., Zhang, W.C., Niu, G.Y., Ren, L.L., Qin, C.Z., 2008. Spatial statistical properties and scale transform analyses on the topographic index derived from DEMs in China. Computers & Geosciences 35, 592602. Zevenbergen, L.W., Thorne, C.R., 1987. Quantitative analysis of land surface topography. Earth Surface Processes and Landforms 12, 4756.

You might also like