| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
a Dep. of Crop Sci., North Carolina State Univ., Box 7620, Raleigh, NC 27695-7620
b Dep. of Crop Sci., North Carolina State Univ., Vernon James Res. and Ext. Cent., 207 Research Rd., Plymouth, NC 27962
* Corresponding author (mflowers{at}cropserv1.cropsci.ncsu.edu)
Received for publication January 5, 2002.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Abbreviations: B, blue (band) • CIR, color infrared • DVI, difference vegetation index • G, green (band) • GS, growth stage • NDVI, normalized difference vegetation index • NIR, near infrared • OSAVI, optimized soil-adjusted vegetation index • R, red (band) • RVI, ratio vegetation index • SAVI, soil-adjusted vegetation index
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Several N management strategies have been developed for increasing tiller density in winter wheat stands in the Southeast. Scharf and Alley (1993) recommended applying N at GS 25 when the average tiller density was <1000 tillers m-2. Weisz et al. (2001) examined optimum N rates across a wide range of GS-25 tiller densities and suggested a critical threshold of 540 tillers m-2. They found that for wheat fields with GS-25 tiller densities below this threshold, optimum grain yields were obtained by applying N at GS 25. Fields with higher tiller densities were best managed by withholding N until GS 30. Therefore, for growers to maximize winter wheat grain yields, they must know the tiller density at GS 25 and be able to apply N quickly if N is required.
To assist growers in estimating GS-25 tiller density, Flowers et al. (2001) developed a remote sensing technique based on NIR digital counts from color infrared (CIR) aerial photographs. However, there are many factors that can influence the measured reflectance in the visible and NIR spectrums and could complicate the use of CIR aerial photographs to estimate GS-25 tiller density across environments. Crop conditions such as weed populations (Menges et al., 1985), plant diseases (Colwell, 1956), insect damage and water stress (Wildman, 1982), varietal differences (Stone et al., 1996; Hatfield, 1990), plant nutrition (Wildman, 1982), and differences in soil backgrounds (Huete et al., 1985; Hatfield, 1990; Bausch, 1993) have all been shown to affect measured reflectance. Atmospheric effects (Jackson et al., 1983), sun angle (Avery and Berlin, 1992), bidirectional reflectance, the lack of radiometric correction, digitization processes, camera settings, film exposure, and film processing may also affect the measured reflectance in CIR aerial photographs. Because of the large number of factors that could potentially influence the relationship between measured reflectance and crop measurements, Blackmer et al. (1996) concluded that ground observations or the inclusion of known reference conditions could benefit the use of remote sensing across environments. Consequently, the NIR remote sensing technique developed by Flowers et al. (2001) used within-field references in the form of a high and low (or bare soil) tiller density measurement to remove differences across environments and estimate GS-25 tiller density. This NIR remote sensing technique was easy to use and required minimal ground truthing. Using the critical tiller threshold (540 tillers m-2) reported by Weisz et al. (2001), the NIR remote sensing technique correctly recommended N at GS 25 82% of the time. However, the NIR technique was established based on data from a limited number of wheat fields under ideal conditions (without weeds, across uniform soil types, and a single wheat variety).
Consequently, our first objective was to validate the NIR remote sensing technique (Flowers et al., 2001) over a wide range of field environments. In addition to testing the method in a large number of fields, we specifically wanted to determine if wheat varieties, tillage practices, soil types, and weed populations interfere with the NIR remote sensing technique's estimation of GS-25 tiller density. In developing the technique, Flowers et al. (2001) found that NIR digital counts were more robust estimators of GS-25 tiller density compared with a wide range of spectral indices derived from CIR aerial photographs. Consequently, our second objective was to confirm that NIR digital counts continued to be the best estimator of GS-25 tiller density across a larger and more varied data set. Ultimately, the adoption of a remote sensing technique in this production system will not be determined by its ability to accurately predict tiller density but rather by its effectiveness in determining if N applications are required at GS 25. In this light, our third objective was to test the accuracy of GS-25 N application recommendations made using this NIR remote sensing technique (Flowers et al., 2001) when combined with the critical GS-25 tiller density threshold for winter wheat reported by Weisz et al. (2001).
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
|
At all sites, GS-25 tiller density was determined at the center of each sample location by averaging two 1-m sections of row within a 1.2-m radius circle. For all sites, weed population was visually estimated at each sample location at the time of GS-25 tiller density measurement. Because a wide range in weed populations did not exist at any site, the weed population data was simplified into two data categories (present or absent) for analysis. Sample locations were considered to have a weed population if weeds were visually estimated to occupy >50% of the surface area at the sample location. A visual estimation of >50% was chosen for simplicity and not based on crop or weed characteristics.
All sites except K-5 contained a single soil type. Due to the size of the K-5 site (approximately 15 ha), there were significant portions of the site in a Lynchburg sandy loam, a Goldsboro loamy sand, and a Norfolk loamy sand. To differentiate soil boundaries, a digitized version of the soil boundaries published by Aull (1972) for the K-5 site was used.
Aerial Photography
Remote sensing was performed as described previously (Flowers et al., 2001). Latitude and longitude for all sample locations and of four aerial targets placed at field corners were determined using a differential global positioning system (DGPS) receiver with 1-m accuracy (Trimble AgGPS 132, Trimble Navigation, Sunnyvale, CA). Aerial photographs were taken from a belly-mounted platform using a 35-mm Canon 81 camera (Canon USA, Lake Success, NY) on the same day (26 Jan. 2001 for K-1, K-2, K-3, K-4, K-5, and K-6) or within 1 wk (26 Jan. 2000 for P-2 and 10 Feb. 2000 for P-1, W-1, and B-1) of tiller density measurements. Color infrared film (Kodak Ektachrome 153) along with a Kodak Wratten gelatin filter number 15 (Eastman Kodak Co., Rochester, NY) were used for the aerial photographs. A series of aerial photographs with differing exposures bracketed on F/8 and a shutter speed of 1/250 s were taken at each site. All CIR film was AR-5–processed to obtain false CIR slides. With the exception of K-5, aerial photographs were taken at altitudes as low as possible while ensuring that each site was contained in a single photograph. This resulted in a range of altitudes (approximately 854 m) across sites. All aerial photographs were taken on cloudless days between 1200 and 1400 h standard time.
Digitization of Images and Photographic Analysis
Photographic analysis and digitization of images were performed on the positive false color slides from the CIR film as described by Flowers et al. (2001). Slides were digitized using the procedure described by Blackmer et al. (1996) with a Konica slide scanner (Konica Q-Scan, Konica Corp., Mahwah, NJ) and the software package Adobe Photoshop v 4.0 (Adobe Syst., San Jose, CA). The brightness and contrast were not adjusted on the digitized image. The digitized image was not sharpened using the scanner software. The image was scanned with a resolution of 47 pixels mm-1, with each pixel representing a range in area of 0.25 to 0.48 m2 of ground area. The range in ground area was due to differences in altitude when the image was taken. With the exception of K-5, each field was contained within a single aerial photograph, and consequently, all comparisons were limited to within a given photograph. The digitized images were rectified in ERDAS Imagine (ERDAS, 1997) using the latitude and longitude of the four aerial targets at each site. Root mean square error (RMSE) was calculated for each rectified image and ranged between 0.1 and 1.5 m.
The K-5 site was 15 ha in size, and two sequential aerial photographs were required to capture the entire site. The sequential aerial photographs were taken on the same pass and at the same altitude. Each aerial photograph covered approximately 60% of the site, resulting in approximately 10% of the site being common to both aerial photographs. For data analysis, the NIR digital counts from the two sequential aerial photographs were combined. A 0.5- by 0.5-m grid was draped over the portion of the site contained within each aerial photograph. The pixel closest to the center of each grid cell was selected. This resulted in 33 193 field locations, each represented by two pixels (one pixel in each aerial photograph). The NIR digital counts for these pixels were determined (see below), and the NIR digital counts from the two photographs were regressed against each other (SAS Inst., 1998). The resultant linear model (r2 = 0.71) was used to convert NIR digital counts for each pixel in one photograph to approximate those in the other. The two photographs were then combined into a single adjusted image.
Spectral Indices
Digital counts representing the spectral reflectance for each sample location were derived using ERDAS Imagine as described by Flowers et al. (2001). Color infrared film emulsions respond to light within the visible and NIR (490–900 nm) regions of the electromagnetic spectrum. The digitized images are represented by 24-bit true color with three bands [8-bit red (R), 8-bit green (G), and 8-bit blue (B)]. At each pixel in the image, the primary color value represents RGB digital counts within the range from 0 to 255. The spectral properties of CIR film result in wide overlapping wavelength bands. In our case, Band 1 (NIR) of the image covered the wavelengths between 490 and 900 nm, Band 2 (R) between 490 and 700 nm, and Band 3 (G) between 490 and 620 nm. While these bands overlap, differences in spectral sensitivity exist between them. Maximum sensitivity occurs at 730 nm in the NIR band, 650 nm in the R band, and 550 nm in the G band. These differences in spectral sensitivity may offer increased information through the use of a spectral index.
In that light, several spectral indices in addition to digital counts for the NIR band were examined. A normalized difference vegetation index (NDVI; Yang and Anderson, 1999) was determined using the digital counts from the NIR and R bands such that:
 |
A normalized NIR value (Jain, 1989) was derived from all bands such that:
 | [2] |
A ratio vegetation index (RVI) (Jordan, 1969) and a difference vegetation index (DVI) (Tucker, 1979) were calculated as:
 | [3] |
 | [4] |
A soil-adjusted vegetation index (SAVI) (Huete, 1988) and an optimized soil-adjusted vegetation index (OSAVI) (Rondeaux et al., 1996) were calculated as:
 | [5] |
 | [6] |
Finally, the digital counts for the NIR and R bands were summed (Wanjura and Hatfield, 1987).
Data Analysis
For P-1, P-2, and K-1, analysis of covariance (general linear models, SAS Inst., 1998) was used to determine if wheat varieties influenced the relationship between GS-25 tiller density and NIR digital counts, NDVI, normalized NIR, RVI, DVI, NIR + R, SAVI, and OSAVI. Variety was considered as a class variable, with GS-25 tiller density as the covariate (linear and quadratic terms). Analysis of covariance was also used at P-2, W-1, and K-5 to determine if weed population and soil type influenced the relationship between GS-25 tiller density and NIR digital counts, NDVI, normalized NIR, RVI, DVI, NIR + R, SAVI, and OSAVI. Weed population and soil type were considered as class variables and GS-25 tiller density as the covariate. Pearson correlations (SAS Inst., 1998) were used to compare NIR digital counts, NDVI, normalized NIR, RVI, DVI, SAVI, OSAVI, and the summed NIR + R with GS-25 tiller density.
Flowers et al. (2001) modified a procedure originally described by Blackmer et al. (1996) and developed a technique for estimating GS-25 tiller density across environments using NIR digital counts and minimal ground truthing. Initially, relative NIR digital counts (NIRRelative) and relative tiller density (TDRelative) are determined for each site such that:
 | [7] |
 | [8] |
 | [9] |
Rearranging Eq. [7–9] results in the following equation for predicting tiller density (TDPredicted) from remotely sensed data at any given site:
 | [10] |
To validate this NIR remote sensing technique for determining TDPredicted, the data from each of the 10 site-years studied in this research were first converted to NIRRelative and TDRelative (Eq. [7] and Eq. [8]) and then regressed as in Eq. [9]. The resulting slope and intercept of the linear regression from each site-year was then compared to that found previously and described in Eq. [9].
In the limited data set used to develop Eq. [10], Flowers et al. (2001) found that it correctly predicted the tiller density to be either above or below the critical tiller density threshold (540 tillers m-2) recommended for making N application decisions (Weisz et al., 2001) 82% of the time. Equation [10] was used to calculate TDPredicted for each sample location in the 10 site-years studied here. A quadrant plot (Flowers et al., 2001) was used to determine the accuracy of these predictions in making GS-25 N recommendations. A plot of TDPredicted vs. TD was divided into quadrants at TDPredicted and TD equal to 540 tillers m-2 and the percentage of data in each quadrant computed. Data in the upper right and lower left quadrants represented instances when N management decisions based on Eq. [10] would have been correct. Data in the upper left and lower right quadrants represented the percentage of incorrect N management decisions based on Eq. [10].
|
RESULTS AND DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
|
|
While differences in the spectral reflectance between wheat varieties exist, these differences may not affect remote sensing applications. Fields are typically sown with a single variety, and as a general practice, aerial photographs typically include a single field. Therefore, remote sensing would be used on a field-by-field basis and thus avoid any effect wheat variety may have on spectral reflectance.
Weed Effect
Italian ryegrass (Lolium multiforum Lam.) was present in W-1 at GS 25 (49 of 106 sample locations). The GS-25 weed density significantly affected all spectral measurements at W-1 (Table 3). The degradation of the relationship between GS-25 tiller density and NIR digital counts due to the presence of Italian ryegrass at W-1 is shown in Fig. 1
. The improvement in linear correlation between all spectral measurements and tiller density by the removal of weedy sample locations (n = 49) at the W-1 site is shown in Table 4. Spectral indices were not any more robust than NIR digital counts when weeds were present at this location (Table 4).
|
|
Of all the problems faced by the remote sensing, weed population may be the easiest to resolve. Good weed control is an essential component of current agricultural practices. Due to the expense of remote sensing, fields with good weed control may be candidates for remote sensing while weedy fields will not be candidates. Therefore, in practice, weedy fields may be avoided.
Soil Type Effect
Bausch (1993) reported that soil background color significantly altered NDVI values throughout the vegetative growth period of corn (Zea mays L.). A similar result was found at K-5 where soil type significantly influenced the relationship between GS-25 tiller density and NIR digital counts (Table 3). When NIR digital counts were regressed against GS-25 tiller density, there was no significant difference in the intercept and slope between data from the Norfolk and Goldsboro loamy sand areas (Table 5). Therefore, the Norfolk and Goldsboro loamy sand sample locations were combined to produce a single linear relationship between NIR digital counts and GS-25 tiller density (Fig. 2
and Table 5). The NIR digital counts for the same tiller densities were lower for the Lynchburg sandy loam compared with the Goldsboro or Norfolk loamy sand. This resulted in one relationship between GS-25 tiller density and NIR digital counts for the Lynchburg sandy loam and a second relationship for the Norfolk and Goldsboro loamy sand.
|
|
Choice of Spectral Measure
Consistent significant linear correlations were found between tiller density and NIR digital counts, normalized NIR, NDVI, RVI, DVI, SAVI, and OSVI in all 10 site-years (Table 4). As reported by Flowers et al. (2001), there appeared to be no advantage in using a spectral index derived using wide-band spectral data from CIR aerial photographs over NIR digital counts to estimate tiller density. Consequently, only NIR digital counts were used for further analysis.
Validation of Remote-Sensed Nitrogen Application Decisions
The intercept, slope, and r2 of NIR digital counts regressed against GS-25 tiller density for all 10 site-years are reported in Table 5. A single equation might be used to predict tiller density across environments if the slope and intercept values were similar for each site. However, consistent with previous findings (Flowers et al., 2001), the slope and intercept values are statistically different across the 10 site-years (Table 5). To remove these differences across environments, Eq. [7] and Eq. [8] were used to convert NIR digital counts and tiller densities to relative values that were regressed against each other. The resultant slope and intercept for 9 of the 10 site-years (Table 6) were not significantly different from those previously reported (Eq. [9]). These nine site-years represented a wide range of environmental conditions, geographical areas, and wheat varieties. Soil types at these sites included a dark mineral organic soil at B-1, lightly colored sandy loams at W-1 and K-1 through K-6, and a red clay soil at P-1 and P-2. Both conventional tillage and no-till with high-residue systems were included. The effectiveness of this technique to remove environmental differences across this range of soils, tillage systems, and varieties indicate that the procedure is robust.
|
Making Nitrogen Application Decisions
Equation [10] was used to determine predicted tiller density (TDPredicted) for data from all 10 site-years (Fig. 3)
. The relationship was linear with a slope of 0.96 and an intercept of -3.82 (not significantly different than a slope of 1.0 and an intercept of 0.0). The NIR remote sensing technique (Eq. [10]) accounted for 76% of the variation (r2) between predicted vs. measured GS-25 tiller density across the 10 site-years.
|
|
|
CONCLUSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Soil type and wheat variety (Table 3) influenced the relationship between tiller density and all spectral measurements. The influence of soil type and wheat variety are significant factors in utilizing remote sensing as they directly influence the amount of actual data that must be collected in the field. To use Eq. [10], the variables TDmax, TDmin, NIRmax, and NIRmin must be known for a given field (or photograph). If the soil type or wheat variety changes across the field, values of all four parameters must be determined for each soil type or wheat variety.
Our second objective was to confirm that spectral indices did not improve the ability to determine GS-25 tiller densities compared with NIR digital counts. Near-infrared digital counts, normalized NIR, NDVI, RVI, DVI, SAVI, and OSAVI were strongly and consistently correlated with GS-25 tiller density (when no weeds were present) across all 10 site-years, six varieties, six soil types, and two tillage systems (Table 4). There was no advantage of using a spectral index derived using wide-band spectral data from CIR aerial photographs compared with NIR digital counts alone. Additionally, we also determined that linear models of spectral indices or NIR digital counts and GS-25 tiller density were not improved by considering quadratic terms (Table 3).
Our final objective was to validate the remote sensing technique (Eq. [10]) for making GS-25 N application decisions. Our research shows that Eq. [10] was highly effective in estimating GS-25 tiller density (n = 978, r2 = 0.76; Fig. 3) across all 10 site-years. Using this NIR remote sensing technique and a critical tiller density threshold for applying N at GS 25 (Weisz et al., 2001), correct application decisions were made 85.5% of the time.
|
ACKNOWLEDGMENTS |
|---|
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  R. P. Sripada, D. C. Farrer, R. Weisz, R. W. Heiniger, and J. G. White Aerial Color Infrared Photography to Optimize In-Season Nitrogen Fertilizer Recommendations in Winter Wheat Agron. J., October 15, 2007; 99(6): 1424 - 1435. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 N. Hong, J. G. White, R. Weisz, M. L. Gumpertz, M. G. Duffera, and D. K. Cassel Groundwater Nitrate Spatial and Temporal Patterns and Correlations: Influence of Natural Controls and Nitrogen Management Vadose Zone J., January 24, 2007; 6(1): 53 - 66. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. P. Sripada, R. W. Heiniger, J. G. White, and A. D. Meijer Aerial Color Infrared Photography for Determining Early In-Season Nitrogen Requirements in Corn Agron. J., June 5, 2006; 98(4): 968 - 977. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
N. Hong, J. G. White, R. Weisz, C. R. Crozier, M. L. Gumpertz, and D. K. Cassel Remote Sensing-Informed Variable-Rate Nitrogen Management of Wheat and Corn: Agronomic and Groundwater Outcomes Agron. J., March 2, 2006; 98(2): 327 - 338. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. P. Sripada, R. W. Heiniger, J. G. White, and R. Weisz Aerial Color Infrared Photography for Determining Late-Season Nitrogen Requirements in Corn Agron. J., September 19, 2005; 97(5): 1443 - 1451. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
N. Hong, J. G. White, M. L. Gumpertz, and R. Weisz Spatial Analysis of Precision Agriculture Treatments in Randomized Complete Blocks: Guidelines for Covariance Model Selection Agron. J., June 17, 2005; 97(4): 1082 - 1096. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
B. L. Ma, K. D. Subedi, and C. Costa Comparison of Crop-Based Indicators with Soil Nitrate Test for Corn Nitrogen Requirement Agron. J., March 1, 2005; 97(2): 462 - 471. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. B. Phillips, D. A. Keahey, J. G. Warren, and G. L. Mullins Estimating Winter Wheat Tiller Density Using Spectral Reflectance Sensors for Early-Spring, Variable-Rate Nitrogen Applications Agron. J., May 1, 2004; 96(3): 591 - 600. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| The SCI Journals | Crop Science | Vadose Zone Journal | |||
| Journal of Natural Resources and Life Sciences Education |
Soil Science Society of America Journal | ||||
| Journal of Plant Registrations | Journal of Environmental Quality |
The Plant Genome | |||
