Determination of Crop Residue Cover Using Field Spectroscopy

Transcription

1 Mission Kearney Foundation of Soil Science: Soil Carbon and California's Terrestrial Ecosystems Final Report: , 1/1/ /31/2006 Determination of Crop Residue Cover Using Field Spectroscopy Susan L. Ustin 1 and Michael L. Whiting 1* Summary The first few centimeters of agricultural soil are the most fertile, holding the greatest amount of soil carbon, and it is the most susceptible to degradation due to rain and wind erosion. Reducing tillage preserves surface crop residues and effectively sequesters organic carbon in soil. Remote sensing methods are suitable for local and regional inventories and monitoring the adoption of conservation tillage practices, however, little work has shown whether readily available satellite imagery is effective for quantifying residue cover. Our assessment of existing and new remote sensing techniques for crop residue cover measurement suggests that the spectral resolutions of hyperspectral and multispectral airborne and satellite sensors can accurately determine the variability of residue and exposed soil. This spatial information will improve carbon sequestration models for cropland management. During spring 2006, the amount of crop residue cover and exposed soil were characterized using digital photography, field spectrometer, and line-transect cover measurements of conservation tillage experimental plots at the West Side Research and Extension Center (WSREC) and the Davis Long Term Research Agriculture Station (LTRAS). An evenly distributed range of mixed crop residue types were collected at 303 sample points within standard, conservation, and no-till tillage plots of grain, corn, cotton, and tomato residues. Counts of point intercepts within digital photographs of the spectrometer field-of-view determined the amount of accumulated residues for pair-wise analysis. These close-range photos provided greater accuracy compared to the "knotted-cord" line-transect method commonly used for estimating cover. The line-transect method over-estimated the cover by 7%. In analysis of four simulated hyperspectral and multispectral bands, the band depth technique Cellulose Absorption Index (CAI) estimated residue within the variation of the two ground measurement methods for entire transects with Residual Standard Errors (RSE) between (where 1.00 is full cover) for all four band combinations. Accuracy was lower for the spectrum fitting technique Angle Near-InfraRed (ANIR) and Shortwave-infrared Angle Slope Index (SASI), and for linear Spectral Mixture Analysis (SMA) analysis with RSE of Objectives This project evaluated existing and new spectral analysis techniques for their potential to determine crop residue cover using hyperspectral airborne and multispectral satellite imagery. The amount of sequestered cropland carbon is related to the amount of residues remaining following tillage. Quantifying the spatial variability in residue amount, not just identifying the tillage practice, is needed along with the soil type, precipitation, and other environment parameters in spatial modeling to refine sequestered carbon predictions. 1 Department of Land, Air and Water Resources, University of California, Davis *Principal Investigator

2 Table 1. Spectral resolution of field spectrometer and convolved spectra simulating airborne and satellite sensors. ASD AVIRIS field spectrometer airborne imaging spectrometer Spectral band width (nm) Number of Bands Surface reflectance, closerange photography, and customary line-transects were used to estimate the amount of cover for a diverse mix of crop residues and bare soils. The surface covers included corn, cotton, and tomato residues from the previous year, with and without recent harvest of small grain stubble, and two soil types in on-going conservation and no-till trials. We tested the efficiency of three remote sensing analysis techniques to detect biogeochemical spectral features: band depth analysis, spectrum shape, and spectral mixture analysis. Point-intercept counts in the photographs were regressed to evaluate the predictive accuracy of four different hyperspectral and multispectral resolutions, characteristic of those used in today's remote sensing instruments. Multispectral Spectral range (nm) Band Ranges Studied (nm) Number of Bands red NIR SWIR1 SWIR2 SWIR3 SWIR4 MODIS satellite imaging spectrometer ASTER satellite imaging radiometer Approach and Procedures Through collaboration with Jeff Mitchell and Anil Shrestha (UC Kearney Agricultural Center), this study measured areal cover of residues and exposed soil with a diverse range of mixed crop residues with line-transect, close-range photo interpretation, and full range field spectrometer. 2

3 Determination of Crop Residue Cover Using Field Spectroscopy Whiting Data Collection Three replicated transects were measured on May 9, 2006, with mixtures of recently harvested rye and wheat, and cotton and tomato residues from the previous year located within seven notill and standard tillage experimental plots at the West Side Research and Extension Center (WSREC), Five Points, CA. Figure 1. Probability distribution of residue cover within conservation, no-till and standard tillage plots. Figure 2. Examples of the digital photos coincident with spectral measurements corn residue in a) standard and b) conservation tillage. On May 26, 2006, data collection was repeated in 10 ridge-till conservation and standard tillage plots with corn and tomato residues at the Long Term Research Agriculture Station (LTRAS), Davis, CA. At 5-foot intervals along the middle line transects, the surface 0.5 m diameter field-of-view (FOV, ~ 1.5 m2) were measured looking downward (nadir) with an ASD (Analytical Spectral Devices Inc, Boulder, Colorado) Field-Pro field spectrometer that measures wavelengths across the visible and solar infrared. Spectra were calibrated using a Spectralon panel (Labsphere, Inc, North Sutton, New Hampshire). The residue within the digital photographs of the spectrometer FOV was quantified with 178 point-intercept counts using 3

4 Determination of Crop Residue Cover Using Field Spectroscopy Whiting SamplePoint software (Booth et al. 2006; USDA ARS, Cheyenne WY, Figure 3. Line intercept of corn and other crops in CT plot. Spectral Processing The mean of multiple spectral measurements were generated and analyzed in Interactive Display Language (IDL, ITT Visual Information Solutions, Boulder, CO). Field spectra and convolved spectra to the spectral resolutions of Airborne Visible InfraRed Imaging Spectrometer (AVIRIS), MODerate-resolution Imaging Spectrometer (MODIS), and Advanced Spaceborne Thermal Emission and Reflection radiometer (ASTER) in table 1. After evaluating a large number of indexes for residue, vegetation, and water determinations, we selected three spectral analysis algorithms to quantify residue and exposed soil: 1) Absorption band depth: Cellulose Absorption Index (CAI = 0.5*(R R2100) - R2100) (Daughtry et al. 2004) estimates absorbed light by cellulose and lignin near 2100 nm. 2) Spectrum shape: Angle Near InfraRed (ANIR) and Shortwave-infrared Angle Slope Index (SASI) algorithm (Khanna et al. 2007) fits the spectrum shape using angles and slopes between red, near-infrared, and SWIR 1 and 2 bands. 3) Spectral mixture analysis: apportions the spectral contributions from mixtures of the pixel's surface components (Ustin et al. 1993). Results The inclusion of standard, conservation and no-till practices provided a full range of residue and soil cover fractions for regression analysis of the spectral data. The range of residue cover measured by photo transects was well dispersed from 0 to 100% cover, figures 1and 2 illustrate the similarity of soil and corn residue in the visible spectrum. 4

5 Line transect Surface corn, small grain, cotton, and tomato residues, and recently sprouted corn, and two soil types, at air dry moisture content, were measured with the knotted-cord line-transect method (USDA-NRCS, 1990) by counting the presence of straw or other residues at each 1-foot marker interval (fig. 3) on three diagonal transects across each plot at 1-m offsets. The line-transect of 50 to 100 counts by three replicates generated 150 to 300 sample counts. A comparison of linetransect estimates to the photo point-intercept measurements showed the line-transect generally over-estimated the amount of cover by 7% (fig. 4). Figure 4. Close similarity of line-transect measurements to photo point intercept. Point-intercept measurements are on 5% lower confidence limit of line transect. Figure 5. Spectral mean (solid line) and 90% confidence limits (dashed) for conservation tillage (black) and standard tillage (green) for a) corn and b) tomato residues. Digital camera samples The photo measurements provided detailed determinations using high sampling rates along the transects and was less influenced by miscounting. Each photo transect measurement estimated the residue and soil with greater than 3,700 points. The photo interpretation method used precise 5

6 marker placement, created no parallax error, and with zoom capability, allowed for detailed repeatable inspection by the interpreter. This system was highly reliable and less susceptible to systematic software errors, although it depended on the skill of the interpreter. Figure 6. Prediction of transect residue cover by the four instrument spectral resolutions. Dashed lines are 95% confidence limits. Figure 7. The fitted spectral shape of dry and moist soil, dry and green vegetation using apex angle in the ANIR triangle and the apex angle and slope of opposite leg in the SASI triangle (Khanna et al. 2007) Spectral estimation of measurements Figure 5 a and b compares spectral measurements of corn and tomato residues in conservation tillage (CT) and standard tillage (ST) made in the trial plots at the LTRAS trials. The strong concave absorption at 1,440, 1,725, and 2,100 nm for corn and tomato residues is indicative of increased concentrations of water, lignin, and cellulose in CT. The reversal of the curve at 2,100 6

7 nm and increased concavity at 2,200 nm for the -OH absorption of secondary clay minerals is observed with increased soil exposure and decreased residue cover in ST. Although there is an overall albedo difference between the mean spectra for CT and ST measurements, the overlap in reflectance at lower cover percentage seen in figure 5 a and b does not allow separation by albedo alone. The dashed lines are the upper and lower 90% confidence limits. The range in the CT spectra indicates that some bare soil was measured within the residue transects. The overlap of the lower confidence limit of the CT spectra with the ST spectra shows a gradient in reflectance, which supports the expectation that a range in residue cover was present for the spectral mixture and specific band analyses. The ability of these spectral analysis techniques to estimate cover, as determined by regressions, is seen in the tables below. The techniques are listed in order of diminishing sensitivity from CAI (mcai, ncai), ANIR/SASI, and SMA for residue sample points and entire transects. Figure 8. Endmember reflectance spectra used to unmix two soil types, grain stubble, corn residues, and green vegetation. Cellulose (2,100 nm) and clay-oh absorptions (2,200 nm) shown with vertical lines. Band depth technique: CAI The ASD spectrometer and hyperspectral AVIRIS imager could use the published band positions (Daughtry et al. 2004), and the algorithm was adjusted to accommodate the band positions of multi-spectral instruments. For ASTER data, the modified CAI (mcai) substituted the 1,650 nm band on the shorter wavelength shoulder for the 2,000 nm band, and the continuum was interpolated to 2,205 nm to measure the 2,165 nm cellulose band depth. The MODIS instrument lacks a band for a shoulder above the cellulose absorption. To compensate, the cellulose absorption was calculated using the following equation between the shorter wavelength 1,640 nm and 2,130 nm as a normalized difference ratio CAI, ncai = (R R 2130 ) / (R R 2130 ). 7

8 Table 2a. Regression coefficients for point sample CAI analysis and photo point-intercept residue cover measurements. Second order regressions CAI ASD AVIRIS ASTER MODIS RSE r^ n Table 2b. Regression coefficients for transect mean CAI analysis. CAI ASD AVIRIS ASTER MODIS RSE r^ n The existing CAI band depth technique provided the best estimator of residue cover at sample points with a Residual Standard Error (RSE) less than 10% residue for all resolutions, except MODIS ncai (table 2a and b). This accuracy is competitive with line transects when compared to the photo point intercepts. ASTER mcai performed nearly as well as the ASD and AVIRIS resolutions, indicating that residue estimation through the cellulose and lignin band depth technique by various band positions around the cellulose band, as long as the cellulose band is present, is robust. While the ASD and AVIRIS hyperspectral data were calculated using the published CAI equation, letting the algorithm select the actual wavelength positions of the maximum shoulders and absorption depth could improve its accuracy. The robustness of using this absorption region is even seen in the more modest accuracy produced by the MODIS normalized difference ratio. Figure 6 illustrates the similarity of the regression predictions over entire transects for all instrument resolutions, all with RSE below 0.10 and r 2 equal to 0.90 or greater. Spectrum shape: ANIR and SASI This is the first known application of this angle and slope technique to determine the amount of residue and soil. Khanna et al. (2007) demonstrated that for classes of land cover types, angle indexes were highly accurate. The ANIR-SASI shape fitting algorithm is explained in figure 7. ANIR and SASI were developed to fit multispectral sensor spectra, and even the variation in band positions of ASTER and MODIS quantifying residue (RSE of 0.17 to 0.18) primarily due to the ability to fit the reflectance change as residue decreases and soil area increases (table 3a through d). The general shape, whether fitted to hyperspectral or multispectral data, is a first order attribute of the surface reflectance that yields consistent accuracy levels among spectral resolutions. While the ANIR-SASI error in point measurements is twice that of CAI, its accuracy 8

9 significantly improved in the transect data indicating high accuracy in large mixed pixels, and thus is a good candidate for additional adjustments. Table 3a. Regression coefficients for point sample ANIR-SASI analysis and photo point-intercept residue cover measurements. Second order regressions ANIR & SASI ASD AVIRIS ASTER MODIS RSE r^ n Table 3b. Regression coefficients for point sample ANIR-SASI analysis and photo point-intercept exposed soil measurements. Second order regressions ANIR & SASI ASD AVIRIS ASTER MODIS RSE r^ n Table 3c. Regression coefficients for transect mean ANIR-SASI analysis and photo pointintercept residue cover measurements. ANIR & SASI ASD AVIRIS ASTER MODIS RSE r^ n Table 3d. Regression coefficients for transect mean ANIR-SASI analysis and photo pointintercept exposed soil measurements. ANIR & SASI ASD AVIRIS ASTER MODIS RSE r^ n

10 Spectral mixture analysis: SMA The linear spectral mixture analysis (SMA) results were constrained by summing the fractions to unity, and the fractions of multiple residue and soil types were consolidated to single "residue" and "soil." SMA can only be performed on the multispectral data by reducing the number of endmembers to no more than the number of bands. Five endmembers were used: two soil types (Panoche clay loam at WSREC and Rincon silty clay loam at LTRAS); bright, fresh grain stubble; flattened darkened corn residues from previous years; and characteristic green vegetation spectrum from cotton (see fig. 8). Table 4a. Regression coefficients for point sample LSU analysis and photo point-intercept residue cover measurements. SMA ASD AVIRIS ASTER MODIS Endmembers RSE r^ n Table 4b. Regression coefficients for point sample SMA analysis and photo point-intercept exposed soil measurements. SMA ASD AVIRIS ASTER MODIS Endmembers RSE r^ n Table 4c. Regression coefficients for transect mean SMA analysis and photo point-intercept residue cover measurements. SMA ASD AVIRIS ASTER MODIS Endmembers RSE r^ n

11 Table 4d. Regression coefficients for transect mean SMA analysis and photo point-intercept exposed soil measurements. SMA ASD AVIRIS ASTER MODIS Endmembers RSE r^ n The soils were significantly different in the visible because of the stronger red reflectance of the WSREC soil (fig. 8). The bright NIR reflectance of the grain and corn residues, and green vegetation easily separates from that of soil. The green vegetation is very distinctive from residues with its strong absorptions in the visible and SWIR regions and strong reflectance in NIR. The minimal amount of green vegetation in the transects precluded a detailed analysis of its contribution. With the consolidation of classes, the green vegetation fraction was dropped since in the photographs and in SMA fractions there was less than 5% green cover, except in three of the 303 scenes, in which it was slightly greater than 10%. SMA is very sensitive to endmember selection, and more site specific than any of the other techniques, as seen by the required inclusion of the two soil types. While the coefficients of determination (r 2 ) were low in the sample points, the generalization over entire transects increased the coefficient significantly, although the error for this method was highest among the techniques (tables 4a through d). Discussion The results of the CAI and ANIR-SASI analysis were not improved with the additional bands in the hyperspectral data beyond those in the multispectral resolutions. This shows the robustness of the techniques with existing multispectral satellite sensors, and practicality of applying the techniques today. Identification of regions in the spectrum that are directly related to important biophysical characteristics was integral in the design of these broad-band instruments. The regressions used are specific to given bands and band regions, however, because these techniques are based on the light absorption by physical properties, they are repeatable or traceable beyond just the statistical inference. The robustness of the CAI was demonstrated with hyperspectral instruments (Daughtry et al. 2004; Nagler et al. 2003), however, the application of this technique to multispectral sensor resolution is new. The utility of ANIR- SASI shape fitting were previously demonstrated in simulated data and in satellite images for classifying land cover (Khanna et al. 2007) and for determining the phenologic stages in crops (Palacios-Orueta et al. submitted). However, this is the first example to demonstrate the accuracy attainable in predicting cover amounts. Accuracy of angle indexes should improve through repeated measurements and identified phenology of associated cropping practices using the available high temporal frequency of MODIS. While Spectral Mixture Analysis was somewhat less accurate in predicting residue cover, by refining the endmembers for the generalized soils and residue we demonstrated that it is possible to obtain high accuracy in quantifying the amount of mixed residues for today s satellite instruments. 11

12 The residue cover and exposed soil estimates were improved by reducing the class types to just residue and soil for spectral analysis, i.e., increasing the samples per class. Also, spatially smoothing the data over entire transects generalized the spectra and reduced band variability. Using the mean spectra for each transect significantly improved nearly all the techniques, except CAI (mcai, ncai). The regressions of the point spectral indexes also required fitting with nonlinear second-order regressions, which indicates an inherit variability that is measured at the fine scale but is eliminated at larger spatial scales. These results are promising in that cover estimates from the satellite scale sensors are of mixed spectra from image pixels 30 x 30 m to 500 x 500 m in size. While the application of close-range photography has been done in the past using dot grid overlays (Daughtry et al. 2004), the SamplePoint software increases the efficiency and accuracy of the count. The amount of residue varies substantially as demonstrated in this study of practices that used the same tillage equipment on the same soil. Using photo point-intercept technique greatly improves the accuracy of research data for model and validation over the linetransects method. The general techniques applied in this study are commonly available and now seen as adaptable to various instrument resolutions and band positions. The techniques promise sufficient accuracy and utility for successful prediction of residue cover content from airborne and satellite image acquisition. This quantification of residue cover now leads to validation of remote sensing data for large spatial areas, and use of this information in understanding soil carbon sequestration, a goal important to the Kearney Mission for understanding carbon cycling with changes in agricultural practices over broad spatial scales. References Booth, D.T., S.E. Cox, and R.D. Berryman Point sampling digital imagery using SamplePoint. Environmental Monitoring and Assessment. 123: , DOI: /s Daughtry, C.S.T., E.R. Hunt, Jr., and J.E. McMurtrey III Assessing crop residue cover using shortwave infrared reflectance. Remote Sensing of Environment 90: Khanna, S., A. Palacios-Orueta, M.l. Whiting, S.L. Ustin, D. Riaño, and J. Litago Development of angle indexes for soil moisture estimation, dry matter detection and landcover discrimination. Remote Sensing of Environment 109, Nagler, P.L., Y. Inoue, E.P. Glenn, A.L. Russ, and C.S.T. Daughtry Cellulose absorption index (CAI) to quantify mixed soil-plant litter scenes. Remote Sensing of Environment 87: Palacios-Orueta, A., M. Garcia, S. Khanna, J. Litago, M.L. Whiting, and S.L. Ustin. (submitted). Summarizing spectral geometry with Spectral Angles Indexes (SAs): Applications to agricultural areas. Remote Sensing of Environment. USDA, Natural Resource Conservation Service The line-transect method of measuring crop residue cover. Technical Note 50. California State Office, Davis, CA. Ustin, S.L., M.O. Smith, and J.B. Adams Remote sensing of ecological processes: A strategy for developing and testing ecological models using spectral mixture analysis. In: 12

13 Scaling Physiological Processes: Leaf to Globe, ed. J. Ehlringer and C. Field, pp: Academic Press. This research was funded by the Kearney Foundation of Soil Science: Soil Carbon and California's Terrestrial Ecosystems, Mission ( The Kearney Foundation is an endowed research program created to encourage and support research in the fields of soil, plant nutrition, and water science within the Division of Agriculture and Natural Resources of the University of California. 13