JEQ Grow Your Career With ASA
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online 9 January 2007
Published in J Environ Qual 36:262-271 (2007)
DOI: 10.2134/jeq2005.0283
© 2007 American Society of Agronomy, Crop Science Society of America, and Soil Science Society of America
677 S. Segoe Rd., Madison, WI 53711 USA
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by D'Acqui, L. P.
Right arrow Articles by Maselli, F.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by D'Acqui, L. P.
Right arrow Articles by Maselli, F.
GeoRef
Right arrow GeoRef Citation
Agricola
Right arrow Articles by D'Acqui, L. P.
Right arrow Articles by Maselli, F.
Related Collections
Right arrow Carbon Sequestration
Right arrow Ecosystem Management
Right arrow Land-use Planning
Right arrow Soil Classification and Mapping
Right arrow Soil Organic Matter

TECHNICAL REPORTS

Organic Compounds in the Environment

Use of Ecosystem Information to Improve Soil Organic Carbon Mapping of a Mediterranean Island

Luigi P. D'Acquia,*, Carolina A. Santia and Fabio Masellib

a Istituto per lo Studio degli Ecosistemi CNR-ISE, Via Madonna del Piano, 50019, Sesto Fiorentino, Italy
b Istituto di Biometereologia IBIMET-CNR, Via Madonna del Piano, 50019, Sesto Fiorentino, Italy

* Corresponding author (dacqui{at}ise.cnr.it)

Received for publication July 19, 2005.

   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Detailed maps of soil C are needed to guide sustainable soil uses and management decisions. The quality of soil C maps of Italian Mediterranean areas may be improved and the sampling density reduced using secondary data related to the nature of the ecosystem. The current study was conducted to determine: (i) the improvements obtainable in mapping soil C over a Mediterranean island by using ecosystem features and (ii) the effect of different sampling densities on the map accuracy. This work relied on field sampling (n = 164) of soil properties measured over the island of Pianosa (Central Italy). Statistical analysis assessing the relationship between soil properties and ecosystem features revealed that the conceptual model of ecosystems defined on the basis of environmental features such as vegetation cover, land use, and soil type was mainly related to the variation of soil organic carbon (OC) content and to the type of Mediterranean environment. The distribution of ecosystems was used to improve the accuracy of soil OC maps obtainable by a simple interpolation approach (ordinary kriging). Substantial improvement was obtained by: (i) stratification into ecosystem types and (ii) applying locally calibrated regressions to satellite imagery that introduced both inter-ecosystem and intra-ecosystem information linked to vegetation features. This study showed that interpolation methods using information on ecosystem distribution can produce accurate maps of soil OC in Mediterranean environments, mostly because of the linkage between soil OC and vegetation types, which are spatially fragmented and heterogeneous.

Abbreviations: AAF, arable fields • HOW, holm-oak wood • NMM, natural Mediterranean Macchia • OC, organic carbon • PPL, permanent pasture land • PW, pinewood • RMSE, root mean square error • ROM, reinvaded olive orchard by Macchia • TM, Thematic Mapper


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
THE anthropogenic impact on land seriously affects the soil ecosystem and in particular is a major factor in soil fertility, with the reduction of soil organic carbon (OC) being the most frequent consequence (Tiessen et al., 1982; Mann, 1986; Schlesinger, 1986; Guo and Gifford, 2002). In this context a key role is played by land cover changes that alter the net production of primary biomass and hence the related organic input to soil. Therefore they can significantly modify biochemical properties of the soil ecosystem. Hence, if and when the pressure of human activities is relieved, vegetation recovery will be strongly influenced by soil conditions (Wardle et al., 1999; Porazinska et al., 2003; Hooper et al., 2005). There is a large variation in the length of time and the rate at which C may accumulate in soil, depending on: (i) the productivity of the recovering vegetation, (ii) the physical and biological conditions in the soil, and (iii) the past history of soil organic C input and physical disturbance (Post and Kwon, 2000). Most of these features are intrinsically associated with specific ecosystems that result from biogeochemical processes acting in a given area. The knowledge of ecosystem dynamic distribution is, therefore, fundamental in evaluating both spatial and temporal variations of soil properties, particularly regarding soil OC content.

Predicting and quantifying the capacity of soil OC sequestration is of specific interest to discern the ecosystem C source/sink relationships. In Mediterranean environments, soils under Macchia (Mediterranean shrubs) and woodland can play an important role as a sink of OC. Detailed maps of soil OC are, therefore, needed to guide correct soil uses and management decisions aimed at increasing soil OC through low cost and environmentally beneficial methods capable of sequestering substantial amounts of atmospheric CO2. These maps are generally produced through the interpolation of point measurements. Different types of sampling approaches have been used to obtain maps based on grid or area sampling including grid cell sampling and zone or direct sampling (Pocknee et al., 1996). However, maps created with traditional survey methods or interpolation of grid soil sampling, although costly, may yield results with inadequate resolution. The result can be poor even if sampling is particularly intensive because an accurate estimate of the spatial discontinuities of soil properties is required to obtain reliable OC maps in heterogeneous environments (Mueller et al., 2001). It is hypothesized that the integration of information from ground measurements and ancillary layers, descriptive of ecosystem distribution, can improve the accuracy of soil OC mapping. Several studies have actually shown that useful predictive relationships exist between quantitative environmental variables and soil properties (McKenzie and Austin, 1993; Odeh et al., 1994; Gessler et al., 1995; McKenzie and Ryan, 1999). More specifically, recent research has demonstrated that significant improvements in the accuracy of a soil C map can be achieved using hybrid mapping techniques that utilize secondary information such as lithology, topography, etc. (Chen et al., 2000; Mueller and Pierce, 2003).

Several statistical procedures can be applied to incorporate environmental features into soil C mapping (Davis, 1973). The performance of various interpolation methods using ancillary data to map soil C has been recently investigated by several authors (Mueller and Pierce, 2003; Simbahan et al., 2006). These studies were usually conducted over agricultural areas where the vegetation cover was relatively homogeneous. The situation is different in seminatural Mediterranean areas that are characterized by extreme fragmentation and heterogeneity of the land surfaces (Lacaze et al., 1996). As already hypothesized, in these cases, information on ecosystem distribution can become essential to achieve accurate soil C maps. The current investigation focused on evaluating the improvement obtained in soil C mapping by different representations of ecosystem distribution in comparison with the cases where such information was not used.

The objectives of this study were: (i) to assess the importance of using ecosystem features to improve the accuracy of soil C mapping in a Mediterranean area and (ii) to evaluate the effect of different sampling densities on the map accuracy. The fulfillment of these objectives is essential in assisting the production of soil C maps for heterogeneous Mediterranean areas. Moreover, the research was expected to cast light on a practical problem of local importance, i.e., accurately mapping the capacity of soil C sequestration of the different ecosystems and of the whole Pianosa Island.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
The investigation was conducted over a flat island with typical Mediterranean features, Pianosa (Tuscan Archipelago, Central Italy), which is a primary study area for investigating C sequestration processes (FLUXNET Project, a global network of micrometeorological tower sites). The study began with the collection of 164 field samples, from which soil properties were analytically determined. These samples were then subjected to a statistical analysis which aimed at assessing the importance of stratification into six ecosystems to discriminate the relationships with the main soil properties of the island. Next, three methodologies were tested to extend soil C over the island. The first was ordinary kriging, which provided the reference interpolation method based only on the use of spatial proximities. The second was kriging of residuals from an ecosystem stratification, which was representative for a common situation where a land cover map can be derived from existing cartography or can be produced by the elaboration of available data. The third method involved locally calibrated regression analyses applied to satellite sensor (Landsat Thematic Mapper) images, which were a straightforward, automatic way to extract and process information related to land surface features.

All tests were performed taking into consideration the number and the related density of sampling points (Atkinson, 1999). While the collection of point samples is generally expensive and time consuming, the local accuracy of all interpolation methods is critically dependent on the density of these samples. This trade-off, which influences the performance of the various methods differently, must be carefully taken into consideration when planning and applying experimental designs over relatively large areas. While, in fact, methods that are purely based on the interpolation of point measurements are particularly sensitive to the sample density, methods that exploit ancillary layers descriptive of spatial land variations (topography, land cover, etc.) are intrinsically more robust to the reduction of sampling points (Davis, 1973).

Study Area
Pianosa Island (42°35'07'' N, 10°04'44'' E) is one of seven islands comprising the Tuscan Archipelago National Park, with a surface area of 10.25 km2 and a coastal perimeter of approximately 20 km. It is characterized by a flat topography (the highest elevation is about 29 m above sea level while the average altitude is about 18 m above sea level) and by a complex mosaic of natural Mediterranean and previously managed ecosystems. The island was settled by a penal colony that managed the agricultural activities for approximately 150 yr but it was abandoned in 1998. The current study started after this period. From 1998 until now few people have been admitted to the island except for study purposes and for maintaining buildings. No agricultural activities have been performed since 1998. However, the effects of agricultural activities on soil and vegetation are still recognizable. Recovering vegetation is gradually spreading over the whole island.

The soils of Pianosa, formed on limestones and sandstones from Pliocene, are characterized by a strong skeletal component, a sandy loamy texture, and a shallow depth. The mean soil depth is less than 30 cm for all ecosystems. Only one soil type is dominant, i.e., the Leptosols group according to the World Reference Base for Soil Resources (ISSS-ISRIC-FAO, 1998).

The original vegetation of Pianosa Island is mainly represented by Mediterranean Macchia, dominated by a mixture of sclerophyllous and deciduous trees, bushes, and grassland. This type of vegetation, which survives predominantly along the coastal perimeter as an association of Crithmo-Limonietea and other endemic species, like Limonium planesiae (Baldini, 2000), has been strongly affected by the agricultural activities of the penal colony.

The Mediterranean vegetation of the island is represented by Rosmarinus officinalis, Cistus spp., and Juniperus Phoenicia, which are typical of calcareous soil (Baldini, 2000) with some Pinus halepensis, Olea europaea, Quercus ilex, and few Eucalyptus Camaldulensis trees along the edge of roads. Abandoned pastures and cultivated fields are dominated by a species association typical of degraded Mediterranean agriculture: Bromus fasciculatus, Daucus carota, Lagurus ovatus, Asphodelus ramosus, Avena barbata, Dactylis glomerata, Plantago lanceolata, Rostraria cristata, Asparagus acutifolius, Petrorhagia saxifraga, and Scabiosa maritima (Colom et al., 2004).

Field Data
An ecosystem type was considered as a conceptual group of sites that are similar in terms of vegetation cover, land use, and type of soil to be grouped together as one map individual. On this basis, six main ecosystem types were identified on Pianosa Island: holm-oak wood (HOW), natural Mediterranean Macchia (NMM), reinvaded olive orchard by Macchia (ROM), pinewood (PW), permanent pasture land (PPL), and arable fields in rotation with pasture and fallow, (AAF).

The main ecosystem boundaries were defined on the basis of aerial photos of the island (source: Regione Toscana Government, Italy) and field surveys and then digitalized in a georeferenced map using ArcView GIS version 3.2 (ESRI, 1999).

Two sets of samples were collected. The first set (n = 114) consisted of a 300-m regular grid to provide a uniform covering of the island and a reasonable number of samples. The second set of samples (n = 50) was collected after the first sampling phase and preliminary analyses. The sampling locations were selected at sites representative of the six identified ecosystems. As a whole, the two samplings produced an average density of 1 sample every 250 by 250 m. Each sampling site was georeferenced using GPS. All points falling in villages or manufactories, roads, and rocks were separately considered.

Since the soils were very shallow and without well-defined horizons, the samples were taken with a Holland auger, sampling the entire soil profile from the surface to the parent material after removal of the litter from the top of the soil. At each intersection grid point three subsamples (one at the referenced point and two within 2 m of distance) were collected and thoroughly mixed to obtain a composite sample.

Total C and N content were determined using a Carlo Erba (Milan, Italy) NA 1500 CHNS Analyzer. Because of the calcareous nature of the soils, the procedure reported by Santi et al. (2006) was used to distinguish organic from inorganic C. Two aliquots of each sample, in two replicates, were analyzed by dry combustion: the first without treatment, to assess the total C and N content and the other after treatment with excess HCl, for carbonate removal, to assess only the organic C content. The analysis of cation exchange capacity (CEC), pH, available P, and particle size distribution were performed according to the SISS methods (SISS, 1985). Sampling depth was given as the average of the depths measured for all points in each ecosystem. Due to the shallowness of the soils the whole depth of the profile (from the surface to the parent material) was sampled. The main physical and chemical characteristics of soil in different ecosystems are given in Table 1.


View this table:
[in this window]
[in a new window]

  		Table 1. Main soil physical and chemical characteristics of different ecosystems (averages of n = 164 samples).

 
Satellite Images
Images taken by the Thematic Mapper (TM) sensor, mounted onboard the Landsat 5 satellite, were used as remotely sensed data. Landsat TM acquires radiation in seven spectral bands (Bands 1, 2, and 3 correspond to blue, green, and red, respectively, Band 4 to near infrared, Bands 5 and 7 to middle infrared, and Band 6 to thermal infrared) with a spatial resolution of 30 by 30 m2 (120 by 120 m2 for Band 6) and a revisiting period of 16 d. Due to these properties, TM data is particularly suitable for mapping and monitoring vegetation features (plant type, density, condition, etc.) on local to regional scales (Horler and Ahern, 1986).

Thematic Mapper images taken on 20 Sept. 2003 were used. The date was chosen to coincide with the soil sampling period conducted at end of summer, which was suitable for characterizing local vegetation features (Hall et al., 1995). These images were completely cloud-free and unaffected by atmospheric perturbations over the island studied.

Statistical Analysis of Field Data
The statistical analysis assessing the relationship between soil properties and characteristics and the conceptual model of ecosystems was accomplished using R software (R Development Core Team, 2004). An analysis of variance (ANOVA) based on the division into ecosystems was performed on the following soil properties and characteristics: CEC, pH, available P, N, OC, inorganic C, particle size distribution (clay, silt, sand), and depth. Goodness of fit was assessed through visual analysis of the studentized residuals.

Extension of Soil Carbon Data
The extension of point measurements over the land surface can be obtained by several statistical techniques, including those based on geostatistics that are among the most efficient and widely applied. Such methods rely on the concept of semivariance analysis, which is an efficient means of characterizing the spatial autocorrelation of georeferenced measurements (Davis, 1973; Pannatier, 1996). Kriging, in particular, is a moving-average method that attempts not only to estimate the value of spatially distributed variables, but also to assess the probable error associated with the estimates. The theory of regionalized variables, from which kriging is derived, has several ramifications permitting the inclusion of external information such as terrain morphology, land cover, etc. The simplest method to use this information is for stratifying the area examined. With particular regard to soil C, any land stratification which is indicative of ecosystem types can be of great use; kriging can then be applied to the residuals of the strata (Davis, 1973).

Other geostatistical techniques are available allowing a more sophisticated consideration of quantitative secondary variables. Regression kriging and cokriging, for example, are capable of using ancillary covariates, such as lithology and morphology, to construct and apply efficient regression models (Mueller and Pierce, 2003; Simbahan et al., 2006).

Similar regression models can be developed using satellite data taken in the optical domain. Some authors have actually applied regression analyses to remotely sensed data to extend measurements of soil organic matter and C over the land surface (Chen et al., 2000; Daniel et al., 2002; Szakács et al., 2004). These studies clearly indicated that indirect relationships exist between soil C concentrations and surface reflectance, mainly linked to the type and quantity of green biomass that is present on the soil (Simbahan et al., 2006).

An efficient method to utilize these relationships is by means of locally calibrated (or weighted) regressions, which are a relatively new family of procedures including some important properties of moving-average operations (Cleveland and Devlin, 1988). They consist of computing a regression model for each estimation point by weighting the training points according to the relevant distances. Locally calibrated regressions can, therefore, be easily applied to image analysis, where regularly distributed measurements of land spectral properties (pixels) are available. For each image pixel a multivariate regression model is calculated by using suitable weights, which give preferential consideration to the nearest training points (Wang, 1990). Consequently, a fundamental step for the application of locally calibrated regressions is the definition of a function to compute these weights. One of the simplest ways is by a negative exponential function of the Euclidean distance. This function is regulated by the distance range, which controls the spatial "fuzziness" (or "smoothness") of the regression models (Maselli, 2002).

The current experiment started with the application of ordinary kriging to extend all soil C point measurements over the surface of Pianosa Island. This required the preliminary definition of a semi-variogram function, which was performed by fitting an exponential model to the experimental semi-variance values (Pannatier, 1996). Next, a jackknifing cross validation strategy was used to evaluate the accuracy obtained. Correlation coefficients (r) and root mean square errors (RMSE) between measured and estimated C values were used as accuracy statistics. These statistics have been used in a similar context by previous authors (Chen et al., 2000; Mueller et al., 2001; Mueller and Pierce, 2003).

The same experiment was performed after a stratification of the sample points in seven classes: six corresponding to the main representative ecosystems and one to roads, villages or manufactories, and rocks. The map of ecosystem types is shown in Fig. 1, with superimposed sampling points. The OC residuals of this map were computed by subtracting the averages of each ecosystem from the original OC values. These residuals were then used to define a new semi-variogram function and apply the kriging estimator as done before. The extended values were finally summed to the OC averages of the relevant strata to recompute the original values. Also in this case a jackknifing cross validation strategy was used to compute the same accuracy statistics as above.


Figure 1
View larger version (30K):
[in this window]
[in a new window]

  		Fig. 1. Ecosystems map of Pianosa with superimposed sampling points of first stage (n = 114) and second stage (n = 50). The classes correspond to the following seven ecosystems: HOW, holm-oak wood; NMM, natural Mediterranean Macchia; ROM, reinvaded olive orchard by Macchia; PW, pinewood; PPL, permanent pasture land; AAF, arable fields; VMR, villages, manufactories, roads and rocks.

 
Regarding the processing of satellite data, the TM images were initially georeferenced by a nearest neighbor resampling algorithm trained on ground control points, obtaining a mean positional accuracy (RMSE) of about one pixel. Then, taking this positional error into account, the soil sampling points were superimposed on the images by using a relocalization procedure aimed at correcting for one pixel displacement between the images and the sampling points. This procedure, which was originally proposed by Halme and Tomppo (2001), operates by looking for local correlation maxima between the remotely sensed data and the land surface variable to extend.

No radiometric and atmospheric correction was applied to the TM images. In plane areas these corrections act as a mere linear transformation of the original data counts, which does not affect the subsequent elaboration using correlation analysis. A false color composite of the study images is shown in Fig. 2.


Figure 2
View larger version (36K):
[in this window]
[in a new window]

  		Fig. 2. False color composite of the Landsat TM images used in the study (RGB = Bands 5, 4, 3). Vegetated surfaces, which are more reflective in Band 4, appear in green, while poorly vegetated or bare surfaces, which are almost uniformly reflective in all three bands, appear in different gray tones.

 
Further processing of the satellite data involved an analysis of the correlations between the original TM bands and the soil OC values. The spectral signatures of the sampling points were extracted from the six optical TM bands (1–5 and 7), and correlation coefficients were computed between these TM bands and soil OC. These correlations served as the basis for selecting the most informative band subset and applying the locally calibrated regression procedure. This application required the definition of the distance range most suitable for the local situation. This was done by applying jackknifing to a series of range values to identify the one that minimized the total RMSE. Then, the extension method was applied to the whole island and its results were again evaluated by using the same accuracy statistics and the same strategy as above.

Finally, all three extension methods (kriging, kriging of ecosystem residuals, and locally calibrated regressions) were repeated with decreasing sampling densities. In particular, starting from the average sampling density, which was of about 1 point every 250 by 250 m2, three degraded densities were considered corresponding to 1 point every 500 by 500, 750 by 750, and 1000 by 1000 m2. In the three cases, subsets of 25, 11, and 6% of the original points were randomly selected for the simulations, repeating them 4, 9, and 16 times, respectively, to consider all available points. The estimates found by jackknifing in each simulation were then summed over the whole original sample to compute accuracy statistics with all points.


   RESULTS AND DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
CONCLUSIONS
 REFERENCES
 
Soil Properties and Conceptual Model of Ecosystems
In the statistical analysis, performed to assess the relationship between soil properties measured in lab and the ecosystem model, frequency analysis of soil OC content in the ecosystems showed a bimodal distribution (not reported) due to some samples (28 of 164 total) whose OC content, >70 g kg–1, was inconsistent with samples from similar ecosystems. For this reason the ANOVA analysis performed on the complete data set was unsatisfactory as shown by the residuals in Fig. 3a (OC). Therefore, the analysis was performed after removing the above-mentioned samples.


Figure 3
View larger version (13K):
[in this window]
[in a new window]

  		Fig. 3. qq-plot of the residuals of the ANOVA analysis for soil organic carbon (OC) using (a) all samples (n = 164), (b) only samples with OC < 70 g kg–1, and (c) all samples for the C/N ratio.

 
The resulting t table for OC content displays three groups (Table 2), and its residuals are shown in Fig. 3b (OC < 70 g kg–1): the first includes NMM, PW, ROM, and HOW, not distinguishable from each other, but well separated from AAF ecosystem (P < 10–3) and less from PPL (P < 0.0498). These results show that the selected environmental features of these ecosystems are well related to the variability of soil OC. The conversion of NMM to AAF or PPL almost invariably resulted in a depletion of native OC stocks. In particular, in the AAF ecosystem soil OC was reduced by about 47% and in PPL by about 16%.


View this table:
[in this window]
[in a new window]

  		Table 2. t table of OC < 70 g kg–1 samples.{dagger}

 
These results are not comparable with others found in Mediterranean environments, due to the lack of similar studies reported in the literature. Studies conducted in other environments reported a decline in OC after land use conversion from forest to crop comparable to that observed here for the AAF ecosystem, and an increase of OC after land use change from forestry to pasture, in contrast to our findings for the PPL ecosystem (Guo and Gifford, 2002; Murty et al., 2002). According to Murty et al. (2002), the conversion of forest to pasture does not, on average, lead to loss of soil OC, although individual sites may lose or gain soil C, depending on specific circumstances, such as application of fertilizer or retention or removal of plant residues. In some environments the decreases in soil OC after pasture establishment can reach 51% (Reiners et al., 1994). The decrease in soil OC after land use change found in the PPL ecosystem can be ascribed to the Mediterranean environments, where pasture can easily promote soil degradation (d'Angelo et al., 2000), and as a consequence, lead to a large soil C loss. For this reason, we expected a larger reduction than that measured in the PPL ecosystem. The unexpectedly higher OC content which was found could be due to the productivity of the recovering vegetation growth on the untended lands that could have increased and maintained the amount of OC, following the findings on soil C accumulation after agricultural abandonment by Knops and Tilman (2000). The two ecosystems ROM and HOW showed values which were not statistically different from NMM. This could be due to the presence of spontaneous Holm oak and olive trees that can be considered part of the natural vegetation of a NMM environment. The conceptual definition of the ROM and HOW ecosystems was based on the larger concentration of these two species in the Macchia vegetation. In this case, the OC content alone is not able to separate the ecosystems from NMM, as revealed by the P values in Table 2. In the PW ecosystem, OC content was not statistically different from NMM. In other similar situations when native forest is cleared for conifer trees, the plantation could have a small effect on soil OC stocks (Guo and Gifford, 2002).

The samples with OC greater than 70 g kg–1 were collected mostly from the NMM, ROM, and HOW ecosystems, but the statistical analysis did not show any significant difference in OC among the ecosystem classes. This may be because only the amount of OC was considered, but not the nature of the organic matter from which it had arisen. In particular, it could be difficult to discriminate OC if decomposition of organic matter was dominant rather than its accumulation. In fact, large differences of soil OC content can be found in the different pools of organic matter (i.e., humified or undecomposed organic material) (Heal et al., 1997). On the basis of these considerations we repeated the statistical analysis with all samples using the C/N ratio (Fig. 3c) from which we can evaluate the dynamics of soil organic matter in the ecosystem (Swift et al., 1979; Knops and Tilman, 2000). In fact in the same ecosystem it is reasonable to assume that similar processes occured thus producing soil organic matter of a comparable nature.

The C/N ratio results are reported in Table 3 and the residuals are shown in Fig. 3c (C/N ratio). Natural Mediterranean Macchia and HOW are not statistically different (P = 0.76), while all the other ecosystems differ from NMM at various levels of significance. The HOW ecosystems can be considered as potentially similar to the Macchia for OC accumulation on Pianosa Island. On the other hand, as discussed above, the Holm oak is an endemic plant of the Mediterranean shrub. In addition, the number of the samples collected from the HOW ecosystem and used for the analysis were not sufficient (n = 12) to give a statistical significance.


View this table:
[in this window]
[in a new window]

  		Table 3. t table of the C/N ratio.{dagger}

 
It was not possible to observe any relationship between the ecosystems and other soil properties and characteristics such as CEC, pH, available P, particle size distribution, and depth. The lack of any relationship within these parameters could be due to various factors: (i) Pianosa soils are homogeneous and flat, and as a consequence the induced processes (natural or due to human activities) of erosion, weathering, etc., are relatively slow; therefore, only small differences in these parameters are observed in soils of various ecosystems. In previous studies regarding different environments, where terrain attributes and soil properties were considered to identify zones of variable soil OC content, the variation of soil properties was found to be mostly related to topography and historical erosion (Terra et al., 2004); (ii) in this type of environment the selected parameters are not capable of identifying ecosystem changes. On the other hand, the decomposition and transformation processes of organic substances in soil are relatively fast compared with those of erosion and weathering because they are promoted by biological activity (microorganisms, roots, and plants). Moreover, the change in vegetative cover strongly affects the nature of the organic input into soil promoting different types of transformation processes. The variation, in a relatively short period of time, of soil OC in this environment could be considered as an index of the changes of the ecosystem linked to land use (Degens, 2001; Guo and Gifford, 2002; Smith, 2005).

The parameters selected for ANOVA seem to be appropriate for the characterization of this type of Mediterranean environment for our purposes. Other parameters could be considered useful to describe the environment to be studied in the conceptual ecosystem model. For example, topography and in particular slope should be important parameters to be considered because of their strong influence on soil properties and in particular on OC distribution and dynamics (Gessler et al., 2000). However, in our case study slope is not considered important since the island is almost flat.

Estimation of Soil Organic Carbon
The experimental semi-variograms, obtained taking into account the original OC data and the OC residuals from the ecosystem stratification, are shown in Fig. 4 together with the fitted exponential models. In both cases, a lag class width of 150 m was used to compute the experimental semi-variances. The configuration chosen for the application of kriging to the original data had a nugget of 1.2 C %2, a sill equal to 6.6%2 C, and a practical range of 190 m (Pannatier, 1996). These values were reduced to a nugget of 1.0 C %2, a sill of 3.5 C %2, and a practical range of 130 m when applying kriging to the residuals from the ecosystem stratification. In both cases the theoretical curves fitted well with the experimental semi-variogram values, producing high correlations and small errors between measured and modeled semi-variances.


Figure 4
View larger version (14K):
[in this window]
[in a new window]

  		Fig. 4. Experimental semi-variances found using the original soil organic carbon (OC) values (squares) and the residuals of these from the ecosystem stratification (triangles), with fitted exponential semi-variograms (r and RMSE indicate the agreement between experimental and fitted semi-variances).

 
The correlations found between OC and surface reflectances in the six optical TM bands are reported in Fig. 5. All correlations are negative, as expected, considering the increased absorption of radiation with increasing vegetation density (Halme and Tomppo, 2001). Maximum correlations were found with Band 3 for the visible interval and with Band 5 for the middle infrared interval. These two bands were, therefore, selected together with Band 4, which is the only one covering the middle infrared spectral range and is generally fundamental to vegetation studies (Horler and Ahern, 1986). Using this three-band subset the optimal configuration for the locally calibrated regression procedure was identified with a spatial range of 360 m.


Figure 5
View larger version (10K):
[in this window]
[in a new window]

  		Fig. 5. Correlations (r) found between soil organic carbon (OC) and digital counts of the six Thematic Mapper (TM) reflective bands. Bands 1, 2, and 3 correspond to blue, green and red, respectively, Band 4 to near infrared, and Bands 5 and 7 to middle infrared.

 
The maps obtained by applying the three extension strategies with the maximum sampling density (1 point ever 250 x 250 m2) are shown in Fig. 6. The relevant accuracy statistics are summarized in Fig. 7. Ordinary kriging performs rather poorly in the extension of the soil OC point measurements in this environment. A rather low r was found (0.407) together with a RMSE higher than 2.5% C. From Fig. 6a it can be noted that such a low accuracy of the method is mainly attributable to its incapacity to correctly reproduce abrupt soil OC variations linked to the presence of different ecosystems. Irregular spatial soil OC variations are in fact present in the map due to different concentrations measured in adjacent points. As expected this indicates that the mere use of spatial autocorrelation is insufficient to correctly model the observed phenomenon. This interpretation is confirmed by the notable improvement in accuracy obtained by the stratification of the ecosystem map. In this case an r of almost 0.6 is reached, with a corresponding relative RMSE decrease of about 10%. Figure 6b shows the relevant effect of the ecosystem distribution on the soil OC map: in this case, abrupt spatial changes are introduced, which enhance the map accuracy. This method, however, is incapable of modeling within-ecosystem soil OC variations linked to different vegetation densities and conditions. This factor is better considered by the locally calibrated regression method, which produces a further improvement in soil OC mapping both in terms of r and RMSE. The introduction of both inter-ecosystem and intra-ecosystem information linked to vegetation types and conditions can be clearly appreciated from the OC map of Fig. 6c.


Figure 6
View larger version (18K):
[in this window]
[in a new window]

  		Fig. 6. Soil organic carbon (OC) maps of Pianosa Island obtained by the three extension methods considered: (a) kriging, (b) kriging of ecosystem residuals, and (c) locally calibrated regressions.

 

Figure 7
View larger version (36K):
[in this window]
[in a new window]

  		Fig. 7. (a) Correlation coefficient (r) and (b) root mean square errors (RMSE) obtained by applying the three extension methods with different sampling densities (D250, D500, D750, and D1000 correspond to average densities of 1 point every 250 x 250, 500 x 500, 750 x 750, and 1000 x 1000 m2, respectively). All trials were performed by a jackknifing approach. All correlations higher than 0.180 are highly significant (P < 0.01).

 
The different sensitivities of the three extension methods to the sampling density can be observed in Fig. 7. The prediction capability of ordinary kriging rapidly degrades to almost insignificant levels for a density of 1 sample every 1000 by 1000 m2. The reason for this behavior is likely the same as that previously mentioned, i.e., a simple interpolation with few sampling points misses soil C spatial variations due to differences in vegetation cover. On the contrary, the decrease in accuracy is less marked when applying kriging of ecosystem residuals, and even less marked when applying locally calibrated regressions to satellite data. In these cases the correlations remain, respectively, higher than 0.5 and 0.6, even with the lowest sampling densities, with relatively moderate increases of RMSE.


   CONCLUSIONS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
REFERENCES
 
The spatial distribution of features such as vegetation cover, land use, and soil type are important to build a conceptual ecosystem model to describe the environment and its link with the variation of OC in soil. The selected ecosystem features were suitable for the area studied, but for other environments they may not be sufficient. Hence, additional or other features could be preferable. The other tested chemical and physical parameters were incapable of identifying ecosystem changes in the environment studied.

The current experiment demonstrated that simple interpolation of point measurements through ordinary kriging was ineffective in accurately mapping soil OC in fragmented Mediterranean areas. The major cause can be attributed to the variability of soil organic matter due to the presence of fragmented and heterogeneous ecosystems typical of this type of environment. In these cases, the accuracy of soil OC maps can be significantly improved by the consideration of ecosystem information through the use of the conceptual model of ecosystems and their distribution. The situation may be different in other, more homogeneous environments, where ordinary kriging was found to produce better performance (Terra et al., 2004).

The investigation also indicated that ecosystem information could be introduced rapidly and efficiently into soil OC maps by the suitable processing of high resolution remotely sensed images. Satellite images taken in the optical domain provide information on the spatial variability of vegetation quantity and conditions existing both between and within different ecosystems (Lacaze et al., 1996). Taking into consideration the fact that the mere vegetation features are indirectly and irregularly linked to soil OC content, a flexible approach is necessary to convey the relevant remotely sensed information into the mapping process. Locally calibrated regressions are an optimal means to reach this objective, since they are capable of taking into account the spatial variability of the relationship between soil OC and land surface spectral properties (Maselli, 2002).

These considerations become even more significant as the sampling density of the ground data decreases. The experiments performed in fact quantified the decreased predictive power of kriging with decreased sampling densities, which became almost null with 1 point per km2. The accuracy of the methods relying on the use of ecosystem features remained relatively high even with very low sampling densities. Such findings have of course fundamental implications for the application of similar techniques to wider Mediterranean areas.

The application of the present results to different environmental situations, however, should be considered with caution, due to the peculiar features of the study island, which has a limited size and shows a high homogeneity in topographic and soil characteristics. The application of similar methodologies to larger areas should be compared with additional spatial variability of environmental factors, such as topography and soil types. Nevertheless, one can assume that this application could benefit from the consideration of relevant ancillary layers (digital elevation models, soil maps, etc.), which were practically useless in the current case.

Regarding remotely sensed data, improvements in mapping accuracy could be expected from the use of multitemporal acquisitions, due to the more abundant information on vegetation cover that would be reached by the consideration of relevant phenological variability.


   ACKNOWLEDGMENTS
 
The work was carried out within research activities of Pianosa_LAB Project. The authors gratefully acknowledge Dr. Ottorino Luca Pantani for his advice regarding statistical analysis and Mr. Dodero Alessandro for his assistance in soil sampling and C/N analysis. Thanks are also due to three anonymous referees whose comments improved the quality of the original manuscript.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 





This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF) Free
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by D'Acqui, L. P.
Right arrow Articles by Maselli, F.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by D'Acqui, L. P.
Right arrow Articles by Maselli, F.
GeoRef
Right arrow GeoRef Citation
Agricola
Right arrow Articles by D'Acqui, L. P.
Right arrow Articles by Maselli, F.
Related Collections
Right arrow Carbon Sequestration
Right arrow Ecosystem Management
Right arrow Land-use Planning
Right arrow Soil Classification and Mapping
Right arrow Soil Organic Matter

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
The SCI Journals Agronomy Journal Crop Science
Journal of Natural Resources
and Life Sciences Education		 Vadose Zone Journal
Soil Science Society of America Journal Journal of Plant Registrations The Plant Genome