Printer Friendly

Mapping surface soil organic carbon for crop fields with remote sensing.

Fertilizers, chemicals, and other crop production inputs are typically applied at uniform rates to an entire field. Because of variations of soil characteristics, including soil organic carbon, it may be desirable to vary production inputs based in part on the soil organic carbon content at various locations within the field. The result of uniform management may be over-application of crop production inputs in some field areas and under-application in others. In addition, uniform applications may also increase the chances of pollution of the environment due to excess application in some field areas. Precision farming technology can be used to change application rates of crop production inputs based on field characteristics (Lowenberg-Deboer and Boehlje, 1996; Rawlins, 1996; Wolf and Buttel, 1996; Joseph, 1998). The soil organic carbon concentrations of surface soil have been used to spatially vary the application rate of some crop production inputs (Blackmer and White, 1998). Surface soil organic carbon concentration affects the activity of many herbicides and the rate of application recommended for some herbicides depends on the soil organic carbon concentration (Hance, 1988). Soil organic C influences plant available nitrogen (Dahnke and Johnson, 1990) and it also affects the soil's ability to adsorb plant nutrients (Havlin et al., 1999). Knowing its concentration may therefore be useful, especially if its spatial distribution could be determined accurately (Blackmer and White, 1998) and for low cost (Wolf and Buttel, 1996; Lu et al., 1997).

In order to determine field variation of soil organic carbon and other properties, different techniques, such as grid sampling and zone sampling, have been used. With grid sampling, eight to ten cores are typically taken for a composite sample to represent a certain size of area, typically from one to three acres. Geographic locations are measured for each composite sample, and maps are created from soil organic carbon lab results and location data using interpolation methods. However, this procedure would miss some high or low areas of soil organic carbon within the composited sample from a grid cell. This problem would be more severe as grid cell size is increased. In addition, grid sampling is labor intensive and expensive both for soil sampling and for analysis. Zone sampling (composite of sampling distinctly different field areas based on factors such as soil type, topographic position or soil color) could be a better method to reduce the number of soil samples, but this method may still miss local areas that are high or low in soil organic carbon.

A third method for mapping soil organic carbon uses remotely sensed imagery, as described by Chen et al. (2000) for a field in Crisp County, Georgia. Remotely sensed imagery, especially aerial photograph image data, can provide data with 2 by 2 m (4.2 ft by 4.2 ft) or higher spatial resolution, which allows the mapping of surface soil organic carbon concentrations at this resolution. In addition, because the remotely sensed imagery can be used to select the areas to sample, the number of soil samples can be reduced. Using this method, sampling areas are selected to cover the range in reflectance, which allows the development of a model to predict the soil organic carbon from reflectance values at any location on the field image. As a result, this method minimizes the number of soil samples needed for organic carbon analysis, which in turn greatly reduces the costs for creating the maps. In the case presented by Chen et al. (2000), the image of the field was an aerial-photograph color slide obtained from U.S. Department of Agriculture (USDA) Farm Service Agency for no cost. The number of soil samples taken was only 10 percent of those needed for sampling on a one-acre grid. Although more time is needed for data processing, the result is a more accurate map for less cost. Since the method was only tested in one field, it needed further study. The objective of this research was to evaluate the method of Chen et al. (2000) over three fields using either an aerial-photograph color slide or a digital aerial-photograph image, to determine if the method has general applicability over a wider range of field conditions.

Methods and Materials

Three crop fields (Rock field, Bowling field, and Adkins field), located in the northwest corner of Crisp County, Georgia, were selected for demonstration of the method. All three fields have been in non-irrigated cotton for the past several years. Their areas were 89, 57, and 112 ac (36, 23, and 45 ha) for the Rock field, Bowling field, and Adkins field respectively. Dominant soils in the Rock field are Norfolk (fine-loamy, kaolinitic, thermic Typic Kandiudults) and Orangeburg (fine-loamy, kaolinitic, thermic Typic Kandiudults). In the Bowling field. major soils are Tifton (fine-loamy, kaolinitic, thermic Plinthic Kandiudults) and Faceville (fine, kaolinitic, thermic Typic Kandiudults). Dominant soils in the Adkins field are Fuquay (loamy, kaolinitic, thermic Arenic Plinthic Kandiudults) and Tifton. These series are typical of the Georgia Coastal Plain. The three fields are gently rolling with slopes generally less than five percent 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 imagery. Low elevation areas, shallow depressions, and other relatively low topographic features are common in all three fields and appear as darker regions on the imagery. These areas are often ponded during intense rainfalls and generally have thicker and darker surface horizons than surrounding higher elevation areas. Duration of ponding is not sufficient to appreciably alter soil morphology, however.

A Spring 2001 aerial photograph color slide of the Rock field with a bare and dry surface was obtained from the USDA Farm Service Agency (Figure 1). In Spring 2000, digital aerial images for the Bowling field (Figure 2) and the Adkins field (Figure 3), with a resolution of 2 by 2 m (4.2 ft by 4.2 ft), were taken by the National Aeronautics Space Administration (NASA) ATLAS sensor. The wavelength of each sensor band is shown in Table 1. Soil samples, numbering 26 for the Rock field, 26 for the Bowling field and 45 for the Adkins field, were obtained in winter 2001, fall 2001, and fall 2000 respectively. Because there was an artificially drawn red line on the color slide, the samples taken from the Rock field were selected from locations far enough from the red line to avoid influence. The soil samples for the Adkins field were randomly divided into two sets. One set with 24 soil samples was used to develop the model. The second set with 21 soil samples was used to check the accuracy of the soil organic carbon map developed for the field. A global positioning system (GPS) system with sub-meter accuracy was used to measure the geographic location for each sample. Sample locations were selected based on the variation in soil color and the apparent surface soil texture, to include the expected range of soil organic carbon levels. The soil sample taken at each location consisted of a composite of nine 2 cm (0.79 in) diameter soil cores taken randomly from the 0 to 15 cm (0 in to 5.9 in) soil depth within a 2 by 2 square m (21.5 square ft) area. The samples were air dried during the next two to three days, sieved with a 2 mm sieve, and stored in plastic containers until analyzed. Total soil organic carbon concentrations were determined with a Leco CNS analyzer (Nelson and Sommers, 1996).

In order to convert the color slide of the Rock field into digital format, it was scanned and geo-referenced into a Universal Transverse Mercator (UTM) projection based on sub-meter GPS measurements of identifiable ground control points. After rectification, the images were re-sampled to 2 by 2 m cell resolution. The rectified image was then converted into ASCII format for further processing (classifying) the image. The Bowling and Adkins fields were already in digital format with 2 by 2 m (4.2 by 4.2 ft) resolution, so this image was directly geo-referenced into UTM based on sub-meter GPS measurements of identifiable ground control points.

In order to reduce the variance among the image pixels caused by micro topography, film processing and scanning, a low-pass filter was applied to each band of the images with a 5 by 5 cell mask. This is an average smoothing filter, as follows:

[P.sub.n](i,j)=[i+2.summation over (k=i-2)][j+2.summation over (1=j-2)] [W(k, 1) X [P.sub.o](k, 1)] (1)

where:

[P.sub.n](i,j) = pixel value for the smoothed image at location (i,j)

[P.sub.o](k,l) = pixel value for the image intensity at location (k,l)

W(k,l) = weight factor for each W with value of 1/N (N = 25 is the total number of pixels within the mask)

range of k = [i-2, i+2], the range of l is [j-2, j+2]

After filtering each image, the next step was to develop the relationship between surface soil organic carbon and image pixel reflectance values. For each field, the pixel intensity values at the sampling locations were determined from the filtered image. Two approaches (Chen et al., 2000) were used to create maps of soil organic carbon for the selected fields. For the first approach, the relationships between surface soil organic carbon concentrations and the pixel intensity values for each band were developed by non-linear regression analysis. Then, the equations for each band were further combined by step-wise linear regression analysis to develop the final regression equation for predicting soil organic carbon of the field. The distributions of soil organic carbon for each field were then derived by applying the final regression equation to each field. The resulting soil organic carbon maps of all three fields were classified with an arbitrary classification.

An alternative approach, a minimum distance-clustering algorithm (Jensen, 1986; Lillesand and Kiefer, 1987), was used to perform a classification to the original image. The bands selected for the classification were the same bands as used in band combination with the step-wise regression analysis that was used in the previous approach. The classified result was further processed to identify the surface soil organic carbon concentration for each class based on the step-wise linear regression results developed in the first approach. The procedure of identifying soil organic carbon for each class was as follows: (i) determine the average, the histogram, and lower and upper boundaries for each class based on the relationship developed with step-wise regression analysis; (ii) combine overlapped classes into one class, and reclassify the result; and (iii) determine the average, lower and upper boundaries for each new class. The final result was further filtered with a majority algorithm to remove "single pixel" classes (noises). The map of surface soil organic carbon for the Adkins field was developed with this second approach. Eight classes were grouped in the map.

For the methods used in this paper to be successfully applied to other fields, a remotely sensed image with bare soil surface is necessary since any crop residue cover would interfere with the reflectance of light from the soil surface. The soil surface should be uniformly dry and tilled. Wet soil or a surface crust in some field areas would introduce more errors (other effects on reflected light). The soil iron concentration might also affect the mapping of soil organic carbon although it did not seem to interfere in our studies.

Results and Discussion

The relationships between surface soil organic carbon concentrations and the pixel intensity values with data ranging from 0 to 255 for each band, were not linear as shown by Chen et al. (2000). Therefore, they were developed by non-linear regression analysis. The equations developed by step-wise regression analysis were different for the three fields because the source image for the Rock field was a color slide and the images for the Bowling and Adkins fields were data from the NASA ATLAS sensor. The image quality from the Atlas sensor varied somewhat by geographic location. Also, the distribution of soil organic carbon differed between fields. The final step-wise regression equations for the three fields were shown as follows:

Rock field: SOC = c1 * exp(-R/c4) + c2 * exp(-G/c5) + c3 * exp(-B/c6) ([R.sup.2] = 0.93) (2)

where:

c1 = -4.1096

c2 = 4.4045

c3 = 3.5291

c4 = 33.5939

c5 = 33.4816

c6 = 92.9493

Bowling field: SOC = c1 + c2 * exp(-B4/c6) + c3 * exp(-B6/c7) - c4 * exp(-B7/c8) + c5 * exp(-B8/c9) ([R.sup.2] = 0.95) (3)

where:

c1 = 0.6683

c2 = 90.69

c3 = 380.49

c4 = -61.98

c5 = 12433.31

c6 = 9.4359

c7 = 5.130

c8 = 8.5345

c9 = 3.1271

Adkins field: SOC = c1 * exp(-B1/c5) + c2 * exp(-B4/c6) + c3 * exp(-B5/c7) + c4 * exp(-B7/c8) ([R.sup.2] = 0.89) (4)

where:

c1 = -10.06

c2 = 14.92

c3 = 16.04

c4 = -817.62

c5 = 29.9759

c6 = 27.6473

c7 = 31.7934

c8 = 7.1974

Where SOC is the percentage of surface soil organic carbon concentration, R, G, B are the image intensity values for the red, green, and blue bands, and B1, B4, B5, B6, B7, and B8 are the image intensity values for ATLAS image bands 1, 4, 5, 6, 7 and 8. c1 to c9 are fitting coefficients with values that differed for each field for the reasons given above. The number of coefficients was also different for each field.

For all three fields, the regression equation was then applied to the pixel intensity values of the image. Then, the maps of surface soil organic carbon concentrations for the three fields were obtained by arbitrarily classifying the results with geographic information system (GIS) software ArcView (Figures 4, 5, and 6). From the maps, we noted some strips with high soil organic carbon values near the boundaries of those fields. These strips were mainly caused by weeds or tree shadows on the field boundaries. A straight line was also displayed in the resulting map (Figure 4) because of an artificially drawn red line on the color slide. The soil samples used for model development and calibration were not within those areas to avoid errors they would cause.

The minimum distance-clustering algorithm was used as an alternative classification method (Jensen, 1986; Lillesand and Kiefer, 1987). This algorithm uses minimum spectral distance to assign a cluster (class) for each candidate pixel. The process begins with an arbitrary number of clusters (classes), and then it processes repetitively until meeting a specified stop condition (or conditions). The results from minimum distance clustering contained a large number of classes; usually two to three times the final number of classes wanted. The classes from minimum distance clustering were further processed to identify the surface soil organic carbon concentrations for each class based on the relationships developed by step-wise regression analysis. For each class, the surface soil organic carbon's for the average, histogram, and the upper and lower boundaries were determined. Based on these values, maps of the surface soil organic carbon were reclassified into the final number of classes. The classes shown in Figure 7 were determined using this approach.

The results of classification would generate many "single pixel" classes, that is, a class in a specific area with only one pixel. These pixels were generally caused by image noise or classification errors. In order to remove the "single pixel" classes, additional data smoothing was done with a majority algorithm. This method sets a pixel value at location (i, j) to the pixel value of the dominant class in the filter mask. The process is as follows: a) choose a suitable mask size (such as 3 by 3, or 5 by 5), systematically move the mask over the image, and do the following at each pixel location (i, j); b) select the pixel class value Pm that has the maximum number (majority) in the mask; and c) reassign the pixel value at location (i, j) to [P.sub.m]. The method could be applied repeatedly with different mask sizes. Figure 7 shows the results with the "single pixel" class removed with a 3 by 3 mask.

The accuracy of the results for the Adkins field was checked for both approaches, based on the second set of soil samples, which were different from the samples used in model development. Both approaches indicated good agreement between the measured and the predicted values with an [r.sup.2] value of 0.81 and 0.87 at P = 0 and a 0.95 confidence level for both approaches. The intercept was not different from 0, so it was not considered in the linear equation (Figures 8 and 9).

Summary and Conclusion

We have described two methods, with arbitrary classes and with image classification, for mapping surface soil organic carbon of three fields using remotely sensed imagery. The image source of one field was a color aerial photography film (a color slide), whereas the NASA ATLAS sensor images provided the image sources for the other two fields. The technology and methodology are simple and accurate enough to be of practical use in agricultural production fields, although it will be necessary to take enough samples to develop a regression equation for each field. When applied to agricultural fields, the method is less expensive and more accurate than traditional methods that employ grid sampling, soil analysis and spatial statistic for developing maps of soil organic carbon.

Compared with grid sampling, the method described for mapping surface soil organic carbon greatly reduced the number of soil samples and therefore reduced the cost of mapping surface soil organic carbon if a free or low cost image with good quality could be obtained. Grid sampling for precision farming is labor intensive and expensive both for soil sampling and for analysis. Based on one-acre grid size, the number of grid samples for the three fields would be 89, 57, and 112. For the method described in this paper, the number of samples was reduced to 26, 26, and 24 which would be 29 percent, 46 percent, and 21 percent of the number required to grid sample at a one-ac grid. A larger field could save more for soil sampling and analysis. The method developed also provided the detailed and accurate description of spatial variation in soil organic carbon. Grid sampling may miss some high or low areas of soil organic carbon within the sampling grid. Even if the individual core samples adequately represent the area sampled, the composite sample will not allow one to describe the variation within the area of the composite sample. Zone sampling has similar problems. For the method described here, the image pixel size was 2 by 2 m, allowing the mapping of surface soil organic carbon concentrations at that spatial resolution.

Further study could be conducted to examine whether an equation relating organic C to image intensity values developed for one field could be applied successfully to nearby fields to describe the spatial variability of their organic C. This should be possible if the fields are similar enough, but further research is needed to evaluate the similarity of fields and test this possibility. If this could be done with good accuracy, the method could be extended to map multiple fields with a single remotely sensed dataset and a single set of soil samples. The cost of data collection and analysis would be further reduced.
Table 1. The wavelength of ATLAS sensor.

Band  Wavelength (um)

 1     0.452 ~  0.468
 2     0.524 ~  0.590
 3     0.585 ~  0.629
 4     0.625 ~  0.682
 5     0.683 ~  0.747
 6     0.755 ~  0.895
 7     1.506 ~  1.758
 8     2.074 ~  2.378
 9     8.186 ~  8.614
10     8.195 ~  8.515
11     8.513 ~  8.953
12     8.991 ~  9.368
13     9.470 ~ 10.196
14    10.261 ~ 11.315
15    11.389 ~ 12.019


Acknowledgements

This work was supported in part by a grant from the U.S. Environmental Protection Agency under provisions of section 319 of the Federal Water Pollution Control Act, as amended in 1987, and by funding from the National Aeronautics Space Administration Space Grant Program.

References Cited

Blackmer A.M. and S.E. White. 1998. Using precision farming technologies to improve management of soil and fertilizer nitrogen. Australian Journal of Agricultural Research 49:555-564.

Chen, E, D.E. Kissel. L.T. West, and W. Adkins. 2000. Field-scale mapping of surface soil organic carbon using remotely sensed imagery. Soil Science Society of America Journal 64:746-753.

Dahnke, W.C. and G.V. Johnson. 1990. Testing soils for available nitrogen. Pp. 127-139. In: R.L. Westerman (ed.) Soil Testing and Plant Analysis. Soil Science Society of America Book Series No. 3. Soil Science Society of America, Madison, Wisconsin.

Hance, R.J. 1988. Adsorption and bioavailability. Pp. 1-19. In: R. Grover (ed.) Environmental Chemistry of Herbicides. Volume 1, CRC Press, Boca Raton. Florida.

Havlin, J.L., J.D. Beaton, S.L. Tisdale, and W.L. Nelson. 1999. Soil fertility and fertilizers: An introduction to nutrient management. Prentice Hall, Upper Saddle River, New Jersey.

Jensen, J.R. 1986. Introductory digital image processing: A remote sensing perspective. Prentice-Hall, Englewood Cliffs, New Jersey.

Joseph, K.B. 1998. Who's minding the farm? GIS World 11(2):46-51.

Lillesand, T.M. and R.W. Kiefer, 1987. Remote sensing and image interpretation. Second edition. John Wiley & Sons, Inc., New York. New York.

Lowerberg-DeBoer, J. and M. Boehlje. 1996. Revolution, evolution or dead-end: Economic perspectives on precision agriculture. Pp. 923-944. In: P.C. Robert, R.H. Rust, and W.E. Larson (eds.) Proceedings of the Third International Conference on Precision Agriculture. Minneapolis, Minnesota, June 23-26, 1996. American Society of Agronomy, Crop Science Society of America, Soil Science Society of America, Madison, Wisconsin.

Lu, Y.C., C. Daughtry, G. Hart, and B. Watkins. 1997. The current state of precision farming. Food Reviews International 13:141-162.

Nelson, D.W. and L.E. Sommers. 1996. Total carbon, organic carbon, and organic matter. Pp. 961-1010. Methods of soil analysis: Chemical and microbiological properties. Soil Science Society of America, Madison, Wisconsin.

Rawlins, S.L. 1996. Moving from precision to prescription farming: the next plateau. Pp. 283-302. In: P.C. Robert. R.H. Rust, and W.E. Larson (eds.) Proceedings of the Third International Conference on Precision Agriculture. Minneapolis, Minnesota, June 23-26, 1996. American Society of Agronomy, Crop Science Society of America, Soil Science Society of America, Madison, Wisconsin.

Wolf, S.A. and F.H. Buttel. 1996. The political economy of precision farming. American Journal of Agricultural Economics 78: 1269-1274.

Feng Chen is a research scientist at the University of Georgia in Athens, Georgia. David E. Kissel is a professor at the University of Georgia in Athens, Georgia. Larry T. West is a professor at the University of Georgia in Athens, Georgia. Doug Rickman is a research scientist at the Global Hydrology and Climate Center for NASA in Huntsville, Alabama. J.C. Luvall is a research scientist at the Global Hydrology and Climate Center for NASA in Huntsville, Alabama. Wayne Adkins is an engineer at the University of Georgia in Athens, Georgia.
COPYRIGHT 2005 Soil & Water Conservation Society
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2005 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Chen, F.; Kissel, D.E.; West, L.T.; Rickman, D.; Luvall, J.C.; Adkins, W.
Publication:Journal of Soil and Water Conservation
Geographic Code:1USA
Date:Jan 1, 2005
Words:3863
Previous Article:Soil organic carbon variability and sampling optimization in Montana dryland wheat fields.
Next Article:Influence of summer cover crops on conservation of soil water and nutrients in a subtropical area.
Topics:

Terms of use | Privacy policy | Copyright © 2021 Farlex, Inc. | Feedback | For webmasters |