SSSAJ Journal of Natural Resources and Life Sciences Education
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online 11 January 2008
Published in Soil Sci Soc Am J 72:186-193 (2008)
DOI: 10.2136/sssaj2007.0028
© 2008 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 Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Chen, F.
Right arrow Articles by Luvall, J. C.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Chen, F.
Right arrow Articles by Luvall, J. C.
Agricola
Right arrow Articles by Chen, F.
Right arrow Articles by Luvall, J. C.
Related Collections
Right arrow Carbon Sequestration
Right arrow Remote Sensing

SOIL & WATER MANAGEMENT & CONSERVATION

Mapping Soil Organic Carbon Concentration for Multiple Fields with Image Similarity Analysis

Feng Chena,*, David E. Kissela, Larry T. Westa, W. Adkinsa, Doug Rickmanb and J. C. Luvallb

a Dep. of Crop and Soil Sciences, Univ. of Georgia, Athens, GA 30602
b Global Hydrology and Climate Center, NASA, Huntsville, AL 35806

* Corresponding author (fchen{at}uga.edu).


   ABSTRACT
TOP
 NOTES
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
Remotely sensed imagery with high spatial resolution has been used to map soil organic carbon (SOC) concentrations at a field scale with greatly increased accuracy and reduced cost compared with grid sampling. The procedure, however, requires each crop field to be sampled and mapped separately. The purpose of this study was to determine if cost could be reduced further by grouping a number of crop fields based on their image similarity, and then mapping them together as one group. Ten crop fields with a bare soil surface were selected from a 2000 NASA ATLAS image. The similarity among these fields was examined with the Ward neural network system (WNNS) using the image histogram features extracted from the image for each field. Seven fields were placed into two groups based on the coefficient of determination (R2) values computed from WNNS, with one group consisting of three fields and the second consisting of four fields. Soil samples were taken from the seven fields along with their global positioning system locations and were divided into two data sets, with one for model development and the other for result checking. Models for mapping SOC concentrations were developed for each group of fields using a single procedure. The resulting maps were checked based on soil sample sets that were not used in model development and showed good agreement between mapped values and lab-determined values, with r2 values of 0.80 for one group of fields and 0.77 for the second group of fields. The models were greatly improved compared with the model developed for all seven fields (R2 was 0.87 and 0.91 for two groups vs. 0.63 for all fields and RMSE was 0.108 and 0.143 vs. 0.219 of SOC percentage). The model developed with similarity grouping was also compared with the model for field-by-field mapping and showed close agreement (R2 was 0.87 for Group 1 vs. 0.89 for Field 2 only in Group 1 and RMSE was 0.108 vs. 0.119 for the same field).

Abbreviations: ANN, artificial neural network • MLP, multilayer perception • SOC, soil organic carbon • WNNS, Ward neural network system


   INTRODUCTION
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
The organic C concentration of surface soil is an important soil property in crop management for guiding agricultural fertilizer and chemical applications (Hance, 1988; Dahnke and Johnson, 1990; Havlin et al., 1999). With the rising concern about global change, soil organic carbon (SOC) may have its greatest influence on environmental processes at a global scale. Topsoil is a large terrestrial reservoir of C because soil is the final destination of the vast majority of photosynthetic C fixed in the earth's ecosystems (Rodriguez-Murillo, 2001; Sparling et al., 2006). Knowing the SOC concentration of soils may therefore be useful, especially if the spatial distribution of SOC could be mapped accurately (Blackmer and White, 1998) and at low cost (Wolf and Buttel, 1996; Lu et al., 1997).

Reflectance in various spectral bands has been correlated with soil properties such as soil organic matter since the late 1960s. The relationships developed between reflectance and soil organic matter were examined for designing spectral sensors (Pitts et al., 1983; Griffis, 1985; Smith et al., 1987; Shonk et al., 1991), and algorithms were developed to transform the output reflectance into the concentration of soil organic matter. Research has shown that the organic matter content can have a linear or curvilinear relationship with reflectance in the visual and infrared range (Baumgardner et al., 1970; Leger et al., 1979; Cihlar et al., 1987; Smith et al., 1987; Sudduth and Hummel, 1988; Shonk et al., 1991; Henderson et al., 1992). Research in recent years has been concerned with predicting SOC at different scales with different data sources, such as remotely sensed imagery (Chen et al., 2000, 2005; Sullivan et al., 2005; Fox and Metla, 2005; Stevens et al., 2006), soil and land use databases (Rodriguez-Murillo, 2001; Brejda et al., 2001; Tang et al., 2006), digital elevation models and their derived attributes (Mueller and Pierce, 2003; Thompson and Kolka, 2005), and measurements with spectroscopic techniques (Ebinger et al., 2003; Cozzolino and Morón, 2006; Stevens et al., 2006). Rodriguez-Murillo (2001) calculated the distribution of SOC in Spain using two soil databases. Brejda et al. (2001) used the National Resource Inventory to estimate the surface SOC content in four major land resource areas of the United States. They also analyzed the effects of land use, hillslope position, and slope aspect on SOC levels, and found that land use was a significant source of variation in all four regions. Mueller and Pierce (2003) introduced terrain attributes such as elevation data as a second data set to improve their SOC map with grid sampling.

Ebinger et al. (2003) applied laser-induced spectroscopy for measuring total soil C in a field. They found that both 193- and 247.8-nm lines were sensitive to C concentration; however, the 247.8-nm line had interference by Fe at 248 nm. McCarty et al. (2002) used diffuse reflectance spectroscopy in the near-infrared and mid-infrared (MIR) regions for measuring soil C in the laboratory. They found that both spectral regions contained information on soil C, and the MIR region contained better information related to soil C. Stevens et al. (2006) used two airborne imaging spectroscopic devices (compact: 405–950 nm; shortwave infrared: 900–2500 nm) to measure SOC contents in heterogeneous agricultural soils. They found that the compact imaging spectroscopy alone could not be used for quantitative prediction of SOC. The combination of the two airborne sensors, however, yielded reasonable results due to its wider spectral range. Chen et al. (2000) proposed a technique for mapping SOC with remotely sensed imagery of bare surface crop fields. The procedure for SOC mapping included image filtering, regression analysis, classification, and reclassification. The technique developed was successfully applied to several crop fields in the Coastal Plain region of Georgia, with each field using between 24 and 30 samples for model development (Chen et al., 2000, 2005). Fox and Sabbagh (2002) proposed a soil-line Euclidean distance technique in their mapping of SOC using red and near-infrared bands of remotely sensed data for two fields. Later, Fox and Metla (2005) compared the three methodologies, including principal component analysis, soil line, and the method proposed by Chen et al. (2000) for three fields with five scenes of imagery. They found that all three methodologies were equal for predicting SOC.

Maps of SOC developed with remote sensing have proven to be both more accurate and less costly than grid-sampling methods (Chen et al., 2000); however, this procedure requires that each crop field be sampled and mapped separately. Therefore, the number of soil samples and costs would still be high for mapping the SOC contents of multiple fields. Upon examining a remotely sensed image with multiple fields, some crop fields look similar in their distribution of image properties such as image color and image texture. If these similarities of image properties also represent similarity of soil properties such as SOC concentration, fields can be grouped based on the image similarity and then SOC concentrations (or other soil properties) can be mapped for a group of fields in a single procedure, thereby further reducing the cost in soil sampling, data analysis, and map processing. The total number of soil samples and the cost of data analysis would be reduced in proportion to the number of fields mapped.

Similarity analysis has been an important operation in image- and video-based information systems. In general, a simple pixel-by-pixel comparison (or exact matching) between the corresponding visual data (e.g., images and videos) is not possible except for highly constrained situations (Lim et al., 2001; Santini and Jain, 1999). Different crop fields are located in different parts of the image and have different sizes (areas) and shapes. To examine their similarity, a variety of features (feature vectors) that describe the characteristics of the visual data would need to be extracted from the visual data (Antani et al., 2002). Examples of the feature vectors could be the color (gray-level or multiple-band) histogram (Antani et al., 2002; Cha and Srihari, 2002; Inoue et al., 2000; Rubner et al., 1998; Zhong and Jain, 2000), texture (Antani et al., 2002; Jain et al., 1999; Manjunath and Ma, 1996; Puzicha et al., 1997; Smith and Chang, 1994), and shape (Del-Bimbo and Pala, 1996; Mokhtarian and Abbasi, 2002; Pala and Santini, 1999; Petrakis and Milios, 1999; Torsello and Hancock, 2004). Similarity among the visual data sets could then be examined by quantifying the relationships between these features with an appropriate criterion such as a distance measure or correlation coefficient (Ma and Manjunath, 1998; Pentland et al., 1994). In the study conducted by Chen et al. (2006), 50 rectangular areas were selected from a 1999 digital orthophotograph to examine their similarity considering different features selected for measuring field similarity. Each rectangular area selected was a subset of a crop field or a forest area. Three types of features (feature vectors), including color histograms, color slopes (representing image local variances), and wavelets (representing image global variances), were extracted from the 50 rectangular areas. Two similarity analysis methods, statistical clustering and artificial neural network, were used to measure the similarity among the rectangular areas. The 50 rectangular areas were then grouped into similarity groups based on the similarity measure. The study indicated that the artificial neural network using the color histogram features gave the best result for grouping similar fields. Further study was needed, however, to apply and test this technique in mapping field properties such as SOC for a group of fields. Consequently, the objectives of this study were to group fields within an image based on the field similarity measured with an artificial neural network using the image histogram features extracted from each field, and then to test how accurate similarity grouping would be for mapping SOC concentrations for a group of crop fields.


   MATERIALS AND METHODS
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
An image with multispectral bands, obtained by the NASA Advanced Thermal and Land Applications Sensor (ATLAS) in the spring of 2000, was used in this study. The NASA ATLAS data included 15 bands from the visible, near infrared, shortwave infrared, and thermal ranges. The wavelength range of each band is given in Table 1 .The thermal bands were not used in this study. The image was georeferenced into the UTM coordinate system based on submeter global positioning system (GPS) measurements of identifiable ground control points and was resampled into 2- by 2-m spatial resolution with eight-bit digital format. Seven crop fields with a bare soil surface (Fig. 1 ) were selected from the image for mapping SOC concentrations. Each field was completely separated from the others by its field boundary. The fields were selected because they were visually similar in a displayed image with a band composition of eight, six, and three. All seven fields are located in the northwest part of Crisp County, Georgia (83°53'20.75''–83°54'16.90'' W, 32°0'49.65''–32°1'54.28'' N). The fields have been in unirrigated cotton (Gossypium hirsutum L.) or peanut (Arachis hypogaea L.) for the past several years, with areas varying from approximately 11 to 29 ha (Table 2 ). The dominant soils for the fields are Faceville (fine, kaolinitic, thermic Typic Kandiudults), Orangeburg (fine-loamy, kaolinitic, thermic Typic Kandiudults), and Tifton (fine-loamy, kaolinitic, thermic Plinthic Kandiudults) series, typical of the Georgia Coastal Plain (Table 2). The fields are gently rolling, with slopes generally <5% and surface textures ranging from loamy sand to sandy loam. The more sandy parts of the fields were generally the lightest colored areas on the images. Low-elevation areas, shallow depressions, and other relatively low topographic features appeared as darker regions on the images. The depressional areas are often ponded during intense rainfalls and generally have thicker and darker surface horizons than surrounding higher elevation areas. The duration of ponding is not sufficient to appreciably alter soil morphology, however.


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

  		Table 1. The wavelength range of the NASA Advanced Thermal and Land Applications Sensor (ATLAS). Refer to www.ghcc.msfc.nasa.gov/precisionag/atlasremote.html (verified 25 Sept. 2007) for other parameters.

 

Figure 1
View larger version (133K):
[in this window]
[in a new window]

  		Fig. 1. The 10 crop fields selected from the NASA Advanced Thermal and Land Applications Sensor (ATLAS) data. Soil samples in circles were used for model development and those in stars were used for model validation. Fields in Group 1 (Flds 1, 2, and 3) are shown with red boundaries and Fields in Group 2 (Flds 4, 5, 6, and 7) are shown with yellow boundaries. The two groups were used for mapping soil organic C concentrations.

 

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

  		Table 2. The basic characteristics of the seven fields, including field size, dominant soil series, number of soil samples obtained from each field, and soil organic carbon (SOC) concentrations from soil analysis.

 
A total of 139 soil samples were obtained from the seven fields, with a range of 13 to 31 samples taken per field (Table 2, Fig. 1). The locations of all soil samples were determined using a GPS system with submeter accuracy. The soil sample taken at each location consisted of a composite of nine 2-cm-diameter soil cores taken randomly from the 0- to 15-cm soil depth within a 2- by 2-m2 area. The samples were air dried during the next 2 to 3 d, ground, and sieved with a 2-mm sieve before analysis. Total SOC concentrations were determined with a LECO (St. Joseph, MI) CNS analyzer (Nelson and Sommers, 1996). Areas sampled were selected to cover the variation in the apparent surface soil texture and the range of soil organic matter levels based on soil color. Three soil samples were not used because they were either too close to the field boundary or under tree shadows.

Similarity analysis for the seven fields was performed using the image histogram features (feature vector) extracted for each field. To examine the similarity (or dissimilarity) performance for fields that are visually dissimilar from the seven fields, three other fields from the image were added for the similarity analysis. The image histogram features (feature vector contains Bands 1–8) were extracted for each field image using a computer program developed for that purpose. Eight histograms (for Bands 1–8, respectively) were extracted for each field in which each single histogram was called a feature subvector. The feature vector extracted from each field was a straightforward concatenation of the feature subvectors that were extracted from different bands. Because the goal of similarity matching was to compute the similarity between visual data sets as a whole, a detailed representation of the image histogram may not improve the result of the similarity analysis but rather may increase the sensitivity to noise and will also result in a larger data volume. The above image histograms were quantified based on an interval value of V in the range [1, 2b] (where b = 8 in our study). The maximum number of distinct components within the image histogram feature vector could be determined using the following parameters: number of bits per band per pixel (b), the value of the image histogram interval (V), and the number of bands (C), as follows:

Formula 1[1]
where Npat is the total number of components for the image histogram feature vector.

An artificial neural network (ANN) model was used for similarity analysis among the feature vectors extracted from the 10 fields. One of the popular ANN architectures is the multilayer perception (MLP) network with backpropagation learning. The architecture of a typical MLP network consists of an input layer, one or more hidden layers, and an output layer of neurons (Fig. 2 ). Input values from the input layer are weighted and passed to the hidden layer(s). Neurons in the hidden layer(s), associated with transfer (activation) functions, "fire" or produce outputs that are based on the sum of weighted values passed to them. The hidden layer(s) pass(es) values to the output layer in the same way. The weights in the network are adjusted through the training process, which is to minimize the sums of square errors between the actual and predicted outputs. The network is learned through the training process by repeatedly examining the input–output relationship and adjusting the model coefficients (i.e., weights) to obtain the best possible agreement between the measured and predicted values (Huang et al., 2004).


Figure 2
View larger version (17K):
[in this window]
[in a new window]

  		Fig. 2. The architecture of a typical multilayer perception network.

 
In this study, the Ward neural network system (Ward System Group, 2001) was used. The Ward neural network system (WNNS) consists of three different MLP networks with backpropagation learning and multiple hidden layers. The data patterns that were imported into the WNNS were the feature vectors (image histogram features) generated by the feature extraction process. The input data set was divided into two data sets, the training data set and the test data set, with a population percentage ratio of 60:40. The training data set was used to train the ANN through a learning process, while the test data set was used to test the goodness of the trained ANN model that was developed. The ANN systems then generated the coefficients of determination (R2) between features of fields to examine the similarity between fields (Fig. 3 ). The value of R2 ranged between 0 and 1, with a high value (close to 1) meaning a high similarity between two fields and a low value (close to 0) meaning a low similarity. The basic procedure for similarity matching with an ANN can be summarized as follows: (i) importing feature vectors into the ANN; (ii) defining the input (independent) and output (dependent) variables; (iii) randomly extracting the training data set and the test data set from all the feature vectors; (iv) designing the ANN architecture (WNNS was applied); (v) training the system and developing the ANN model; and (vi) computing the similarity measure values such as R2. The R2 between the features of different field images were computed from the ANN model and were used as the similarity measure of the fields.


Figure 3
View larger version (12K):
[in this window]
[in a new window]

  		Fig. 3. Input, output, and the architecture of the Ward neural network system (WNNS). The input and output are part of the algorithm structure. In the WNNS, PP is the preprocessor, BP1 to BP3 are the three multilayer perception networks, and RC is the result combination.

 
Based on the R2 from the WNNS, the seven fields were placed into two groups for separate mapping of SOC concentrations (Table 3 ). One group (Group 1) included Fields 1 to 3, and the second (Group 2) included Fields 4 to 7 (Fig. 1). The number of total soil samples was 63 for Group 1 and 73 for Group 2. The soil samples in each group were randomly divided into two sets, ensuring that both sets covered the range of SOC concentrations. Each set consisted of about half of the total number of soil samples; one set was used to develop the model for mapping SOC concentrations (31 for Group 1 and 36 for Group 2) and the second set was used to check the accuracy of the maps of SOC concentration developed for each group. Maps of SOC concentrations for each group of fields were developed with a single procedure (Chen et al., 2000), and the accuracy of the maps was examined with the second set of soil samples not used for model development. The basic steps of SOC mapping were as follows: (i) georeferencing the image (if needed); (ii) filtering the image to reduce or remove image noise that interferes with SOC mapping; (iii) developing the relationship between SOC concentrations and image pixel intensity values; (iv) mapping SOC concentrations with the developed relationship and reclassification; and (v) post-processing (e.g., removing single-class pixels and clipping out areas outside the mapped field).


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

  		Table 3. The values of the coefficients of determination (R2) computed with the Ward neural network system. Two groups of fields (represented with a thick black box and a thick dashed black box) were generated.

 
The relationship between SOC concentrations and image pixel intensity values using all seven fields was also examined to check if the similarity grouping could improve mapping of SOC concentrations. Thirty-four soil samples were randomly selected from the seven fields considering the distribution of SOC concentrations, and the model was developed with stepwise regression. Among the seven fields, Field 2 had been previous mapped with a field-by-field procedure (Chen et al., 2005). Therefore, the map of SOC concentrations using the field-by-field procedure was compared with the map using the similarity group that included Field 2. Thirty-two locations were randomly selected and the predicted values of SOC concentrations for those locations using both approaches (field by field vs. similarity grouping) were examined with a linear regression analysis.


   RESULTS AND DISCUSSION
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
CONCLUSIONS
 REFERENCES
 
Extraction of Image Histogram Features
Three factors were considered in extraction of the image histogram features of each field. First, the areas occupied by terraces and waterways or covered by trees and tree shadows within the field images were masked. Areas with terraces and waterways were not cropped, whereas areas under trees and tree shadows did not truly represent the actual surface information. The pixel intensity values in these areas were basically the noise that disturbed field similarity analysis. Second, the interval value V of the image histogram (in Eq. [1]) was selected. A small value of V would introduce interference for similarity analysis, while a large value of V would suppress the useful information. The determination of V was based on the consideration that the R2 value between two obviously similar fields should be as high as possible, while the R2 value between two obviously dissimilar fields should be as low as possible. The value of V = 4 was selected after examining various V values, and the image histogram features for the fields were computed with a specially developed computer program. Third, the image histogram features extracted from the fields were entered into a data table in which each column was the image histogram of a field image and each row was the number of pixels at a specific interval of pixel intensity values in a specific band for different fields (Table 4 ). We could find that some rows in the data table contained only zero values for all fields (upper part of Table 4). This means that there was no pixel at that specific interval of pixel intensity values. Therefore, these rows (with zero values for all fields) were removed before the data table was used for similarity analysis (lower part of Table 4).


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

  		Table 4. A subset of the image histogram features (hf) for the 10 crop fields. Histogram features for Fields 1 to 10 are given by hf1 through hf10. Each column is the image histogram of a field image and each row is the number of pixels at a specific interval of pixel intensity values in a specific band.

 
Similarity Grouping with the Artificial Neural Network
The input features (image histograms) in the WNNS were examined by three different MLP networks with different activation functions applied to the hidden layer slabs (i.e., groups of neurons in the hidden layers) as the input patterns were propagated and processed through the network. The output from the network combined the results from the three MLP networks to produce a better result than that from a single MLP network. The WNNS could also determine the optimal number of hidden neurons for each hidden layer based on the number of input variables, the number of output variables, and the size of the training set. Different values of the initial weights, learning rates, and other parameters were examined for the WNNS; however, the results did not alter the similarity grouping although the R2 values between feature vectors of fields were slightly different as values of the initial weights, learning rates, and other parameters were varied. A table containing the R2 values between any two fields was created through the WNNS (Table 3).

With the computed R2 values from WNNS, two groups of fields were formed (Table 3, Fig. 1). Group 1 consisted of three fields (Fields 1–3), and Group 2 consisted of four fields (Fields 4–7). The R2 values computed from WNNS between any two fields in Group 1 were 0.68 or higher and in Group 2 were about 0.5 or higher (Table 3). By visually checking, it was determined that Fields 8, 9, and 10 should not be grouped into the above seven fields. The similarity analysis with ANN further proved that neither were they placed into the above two groups nor did they consist of a new group because of the small R2 values. These three fields (Fields 8, 9, and 10) were not considered for further SOC mapping in the study.

Mapping Soil Organic Carbon Concentration for Each Group of Fields
Maps of SOC concentrations were developed for fields in Groups 1 and 2 in which each group was mapped with a single procedure. The relationship between SOC concentration and image pixel intensity values of various bands was developed with stepwise regression analysis using a portion of the soil samples, as noted above. The regression equations and R2 and RMSE values for the two groups of fields are listed in Table 5 .


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

  		Table 5. Regression models for surface soil organic carbon (SOC, %) developed with different approaches. Group 1 includes Fields 1, 2, and 3; Group 2 includes Fields 4, 5, 6, and 7; B1, B3, B4, B5, and B7 are the image intensity values for ATLAS image bands 1, 3, 4, 5, and 7, respectively.

 
The maps of SOC concentrations developed based on the regression equations for the two groups of fields are shown in Fig. 4 . The accuracy of the maps developed for the two groups of fields was checked using the second set of soil samples, those not used in map development (32 for Group 1 and 37 for Group 2). The results of the accuracy check with linear regression analysis for the two groups of fields indicated good agreement between the measured and predicted values (Fig. 5 ), with r2 values of 0.80 for Group 1 and 0.77 for Group 2 at P < 0.0001 and a 95% confidence level. The intercepts for both linear regression equations were not significantly different from zero ({alpha} = 0.05), so they were not considered in the linear regression.


Figure 4
View larger version (59K):
[in this window]
[in a new window]

  		Fig. 4. Maps of soil organic C (SOC) concentrations for the seven fields sampled. The points in Field 2 were locations selected for examining the consistence between the maps developed with the field-by-field approach and the approach using similarity analysis.

 

Figure 5
View larger version (12K):
[in this window]
[in a new window]

  		Fig. 5. The linear relationships between measured and predicted soil organic C (SOC) concentration for the two groups of fields. The dash lines are 1:1.

 
Comparison of Different Approaches
We also examined the relationship between SOC concentrations and image pixel values using all seven fields. The regression equations for all fields vs. the models for Groups 1 and 2 are given in Table 5. The regression equations developed with similarity analysis had better R2 and RMSE values. The R2 values increased from 0.63 with all fields to 0.87 for Group 1 and 0.91 for Group 2, and RMSE values decreased from 0.219 for all fields to 0.108 for Group 1 and 0.143 for Group 2.

Soil organic C concentration for Field 2 had been mapped with the field-by-field approach in a previous study (Chen et al., 2005). The regression equation for the field is listed in Table 5. Compared with the approach using similarity grouping, models developed with these two approaches had close values of R2 (0.87 vs. 0.89) and RMSE (0.108 vs. 0.119). The predicted SOC concentrations using these two approaches were also examined with linear regression analysis to check their consistence. Thirty-two locations from Field 2 were randomly selected (Fig. 4). The predicted SOC concentrations at these locations were extracted from the maps developed using both approaches. A linear regression analysis between predicted SOC concentrations of the two approaches showed a high consistence between the maps developed with the field-by-field approach and the approach using similarity grouping (Fig. 6 ). The r2 values were 0.94 at P < 0.00001 with a 95% confidence level.


Figure 6
View larger version (11K):
[in this window]
[in a new window]

  		Fig. 6. The linear relationship between predicted soil organic C (SOC) concentrations with the field-by-field approach and the approach using similarity analysis for Field 2. The dash line is 1:1.

 

   CONCLUSIONS
TOP
 NOTES
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
REFERENCES
 
In this study, similarity analysis of fields was investigated for crop fields using remotely sensed imagery. Image histogram features were extracted from each field image. Similarity of the fields was studied with the WNNS using the image histogram features extracted from those fields. Seven fields were placed into two groups based on the similarity measured with the R2, which was computed from the WNNS. Results showed that using the image histogram features with the WNNS was effective in determining the similarity of fields in the study.

Maps of SOC concentrations were developed for the two groups of fields; each group of fields was mapped with a single procedure. The resulting maps of SOC concentrations developed for the two groups of fields indicated that multiple fields could be grouped and mapped in a single procedure with good accuracy. Compared with the procedure for single-field mapping of SOC concentrations (Chen et al., 2005), the technique developed in this study could greatly reduce the number of soil samples necessary while still creating maps of good accuracy, therefore greatly reducing the cost of field work and laboratory analysis. For example, if SOC concentrations were mapped for the three fields in Group 1 separately, around 30 soil samples would be needed for each field to develop a good regression relationship. With this assumption, a total of approximately 90 soil samples would be needed for the mapping process. When the three fields were grouped based on their similarity (as in this study), however, only about 30 soil samples were needed for mapping the SOC concentrations of the three fields, therefore saving in soil sampling, laboratory analysis, and mapping.


   NOTES
TOP
 NOTES
ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS AND DISCUSSION
 CONCLUSIONS
 REFERENCES
 
All rights reserved. No part of this periodical may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Permission for printing and for reprinting the material contained herein has been obtained by the publisher.

Received for publication January 18, 2007.


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




This article has been cited by other articles:


Home page
 Soil Sci.Home page
C. Wu, J. Wu, Y. Luo, L. Zhang, and S. D. DeGloria
Spatial Prediction of Soil Organic Matter Content Using Cokriging with Remotely Sensed Data
Soil Sci. Soc. Am. J., May 13, 2009; 73(4): 1202 - 1208.
[Abstract] [Full Text] [PDF]

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 Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Chen, F.
Right arrow Articles by Luvall, J. C.
Right arrow Search for Related Content
PubMed
Right arrow Articles by Chen, F.
Right arrow Articles by Luvall, J. C.
Agricola
Right arrow Articles by Chen, F.
Right arrow Articles by Luvall, J. C.
Related Collections
Right arrow Carbon Sequestration
Right arrow Remote Sensing

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
Journal of Plant Registrations Journal of
Environmental Quality		 The Plant Genome