Published in Agron J 100:320-327 (2008)
DOI: 10.2134/agrojnl2007.0140
© 2008 American Society of Agronomy
677 S. Segoe Rd., Madison, WI 53711 USA
REMOTE SENSING
Estimating Ground Cover of Field Crops Using Medium-Resolution Multispectral Satellite Imagery
Stephan J. Maas* and
Nithya Rajan
Dep. of Plant and Soil Science, Texas Tech Univ., Lubbock, TX 79409
* Corresponding author (stephen.maas{at}ttu.edu).
 |
ABSTRACT
|
|---|
Remote sensing is useful for estimating plant canopy characteristics, such as leaf area index (LAI) and ground cover (GC). When the source of remote sensing data is medium-resolution satellite imagery, plant canopy characteristics can be estimated for numerous fields within an agricultural region. In this study, a procedure was developed to estimate GC of field crops from medium-resolution satellite image data in the red and near-infrared (NIR) spectral bands. In the procedure, GC is estimated from the ratio of the perpendicular vegetation index (PVI) value calculated for an image pixel to the PVI value corresponding to full vegetation canopy. Two main advantages of this procedure are that it does not rely on empirical relationships, and that it can use raw satellite digital count data without conversion to surface reflectance or normalization for scene-to-scene differences in brightness. A field study was conducted in 2006 in the Texas High Plains to collect ground-based observations of GC for 31 agricultural fields containing various crops for testing the procedure. The GC for these fields was estimated using the procedure from Landsat-5 TM image data acquired on four dates during the growing season. Statistical analysis of the linear regression between satellite-based estimates of GC and corresponding ground-based observations of GC indicated that the regression was not different from a 1:1 relationship. Statistical analysis also indicated that the average of the satellite-based estimates of GC was not significantly different from the average of the ground-based observations of GC. The results suggest that, on average, estimates of GC determined using this procedure should be within 6% of their true values. The relative simplicity of this procedure should facilitate the quantification of vegetation resources in agricultural regions.
Abbreviations: AAE, average absolute error • CRP, Conservation Reserve Program • DC, digital count • FAPAR, fractional absorbed photosynthetically active radiation • GC, ground cover • LAI, leaf area index • LMM, Linear Mixture Model • NIR, near-infrared • PVI, perpendicular vegetation index • TM, Thematic Mapper • VI, vegetation index
|
NOTES
|
|---|
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.
Received for publication April 15, 2007.
|
INTRODUCTION
|
|---|
REMOTE SENSING has been recognized as an effective means for estimating various plant canopy characteristics, such as LAI, GC, aboveground biomass, and fractional absorbed photosynthetically active radiation (FAPAR), of importance in agricultural and natural resource applications (Asrar et al., 1989; Pinter et al., 2003; Liang, 2004; Hatfield et al., 2004). Typically, the relationship determined between the plant canopy characteristic and remote sensing is empirical, that is, ground-based observations of the plant canopy characteristic are regressed against corresponding remotely sensed observations of the plant canopy, usually expressed as a vegetation index (VI) calculated from reflectance or digital count (DC) observations in various wavelengths (Wiegand et al., 1990; Richardson et al., 1992; Wiegand et al., 1992). In this way, a mathematical relationship is produced for estimating the plant canopy characteristic from remote sensing observations. Medium-resolution imagery from satellites such as Landsat, SPOT, and IRS can provide multispectral remote sensing data for numerous fields within an agricultural region. Applying these data to the mathematical relationships described above would allow the estimation of plant canopy characteristics for these fields, which could serve as the basis for regional assessments of vegetation resources.
Two things hinder the estimation of plant canopy characteristics in this manner. First, there is no guarantee that the empirical mathematical relationships developed between observed plant canopy characteristics and remotely sensed data will be accurate for situations other than those that produced the data for developing the relationships. Wiegand et al. (1992) showed that empirical relationships developed between ground-based observations of plant canopy characteristics (such as LAI) and corresponding values of vegetation indices for five different geographical locations were qualitatively similar, but differed in the specific values of the coefficients in the relationships among the locations. Second, the raw image data from medium-resolution satellites (usually in the form of DC) are affected by the electronic characteristics (gain and offset) of the particular imaging sensor, the solar irradiance (a function of time of day and year), and the clarity of the atmosphere at the time of the observation. Thus, the remote sensing data would need to be converted to some absolute measurement (such as surface reflectance) or normalized to a reference image data set before being used in the empirical relationships to estimate plant canopy characteristics. While procedures have been developed to convert or normalize satellite image data to compensate for image-to-image differences (Mahiny and Turner, 2007), they increase the complexity of the overall process of estimating plant canopy characteristics from medium-resolution satellite imagery.
In this study, we developed a procedure for estimating a plant canopy characteristic (ground cover) from medium-resolution multispectral satellite imagery that does not rely on empirical mathematical relationships and does not require conversion or normalization of the satellite image data. Ground cover was selected as the subject of this study because of our interest in its potential application in estimating crop water use (Maas et al., 2005; Rajan and Maas, 2006). In this article, we describe the theoretical background of the procedure, and present results of testing the procedure against actual field observations of GC.
|
THEORY
|
|---|
When the DC values in the NIR spectral band are plotted vs. the corresponding DC values in the red spectral band for pixels comprising a medium-resolution multispectral satellite image of an agricultural region, a characteristic distribution of points is produced (Jensen, 2000, p. 343; Liang, 2004, p. 249). An example is presented in Fig. 1a
. This roughly triangular distribution of points results from the reflectance characteristics of vegetation and soil, and the mixing of these characteristics within image pixels. Image pixels containing only bare soil lie along the diagonal base of the triangle, forming a line commonly called the "bare soil line." This line is shown in Fig. 1b, which provides a diagrammatic representation of the distribution of points in Fig. 1a. Point a represents the brightest soil in the satellite image, while point b represents the darkest soil in the satellite image. For a single soil type, point a might correspond to pixels containing only dry soil, while point b might correspond to pixels containing only wet soil. Points along the soil line between a and b would correspond to pixels with intermediate levels of soil wetness, or pixels with varying mixtures of wet and dry soil. Points in the distribution above and to the left of the bare soil line represent pixels containing living vegetation. Since living vegetation strongly absorbs light in the red spectral band and strongly reflects light in the NIR band, increasing the amount of vegetation in a pixel decreases the brightness of the pixel in the red spectral band and increases the brightness of the pixel in the NIR spectral band. In the simplest case, this process can be described mathematically by the Linear Mixture Model (LMM),
 | [1] |
 | [2] |
in which DCPixel,Red and DCPixel,NIR are the DC values for a pixel in the red and NIR spectral bands, DCFC,Red and DCFC,NIR are the DC values in the red and NIR spectral bands corresponding to a pixel containing full canopy (GC = 1), and DCSoil,Red and DCSoil,NIR are the DC values in the red and NIR spectral bands corresponding to a pixel containing only bare soil (Heilman et al., 1982; Maas, 1998, 2000). In Eq. [1] and [2], DCSoil,Red and DCSoil,NIR represent the "soil background" brightness, and their values depend on the intrinsic brightness of the soil on which the vegetation is growing (i.e., where the soil would lie along the bare soil line). At full canopy, the vegetation completely obscures the soil surface, so DCFC,Red and DCFC,NIR would represent a single point (point c) in Fig. 1b. Since a dense leaf canopy is capable of absorbing almost all of the light in the red spectral band falling on it (Maas, 1997), DCFC,Red would normally represent a lower limit (the vertical line in Fig. 1b passing through point c) for DCPixel,Red for pixels containing vegetation.
View larger version (18K):
[in this window]
[in a new window]
|
Fig. 1. Pixel digital count (DC) values in the near-infrared (NIR) spectral band plotted vs. corresponding DC values in the red spectral band for a portion of a Landsat-5 image of an agricultural region. (A) Actual distribution of DC values; (B) diagrammatic representation of features of the distribution of DC values; (C) features of the distribution of DC values related to estimating the GC of the image pixel represented by point e.
|
|
Starting with a bare soil pixel with a brightness represented by point a in Fig. 1b, adding vegetation cover would result in a decrease in the pixel DC value in the red spectral band and a concurrent increase in the pixel DC value in the NIR spectral band until full canopy was reached, whereupon the brightness of the pixel would be represented by point c. According to Eq. [1] and [2], this transition would be represented by a straight line connecting points a and c. A similar situation would occur if we started with a bare soil pixel with a brightness represented by point b in Fig. 1b. In this case, the transition described by Eq. [1] and [2] would be represented by a straight line connecting points b and c. However, the boundaries of the triangular distribution of points in Fig. 1a suggest that these transitions are not straight lines, but are curved lines like those shown diagrammatically in Fig. 1b. This apparent discrepancy can be explained by recognizing that the LMM represented by Eq. [1] and [2] assumes that the vegetation and soil components in the pixel are both sunlit. In reality, sunlit plants cast shadows on the adjacent soil surface. Therefore, a more realistic LMM might be,
 | [3] |
 | [4] |
in which DCSunlit-Soil,Red and DCSunlit-Soil,NIR are the DC values in the red and NIR spectral bands corresponding to the brightness of sunlit bare soil, DCShade-Soil,Red and DCShade-Soil,NIR are the DC values in the red and NIR spectral bands corresponding to the apparent brightness of shaded bare soil, and β is the fraction of the soil surface exposed between plants that is sunlit. The apparent brightness of shaded soil is typically less than the brightness of sunlit soil or sunlit leaf canopy in both the red and NIR spectral bands (Maas, 1998). Therefore, DCPixel,Red calculated using Eq. [3] will be less than DCPixel,Red calculated using Eq. [1], and DCPixel,NIR calculated using Eq. [4] will be less than DCPixel,NIR calculated using Eq. [2], for a given value of GC between 0 and 1 and a value of β < 1. The difference between the corresponding calculations will be greatest for the midrange of GC values, and will approach zero as the value of GC approaches either zero or one. For dark soil backgrounds, the effect of shadows on the soil surface between plants can cause the pixel DC value to fall below the red absorption lower limit for certain values of GC, as shown for the curve between points b and c in Fig. 1b.
The leaves of many plants absorb only a small fraction of light in the NIR spectral band, and reflect and transmit the remainder in roughly equal proportions (Gausman, 1985). As a result, the DC value in the NIR spectral band for a pixel with full canopy can continue to increase as the density of the leaf canopy increases, due to the upward scattering and transmission of light by leaves lower in the plant canopy. Thus, it is possible to observe a spike of points in the plot of NIR vs. red pixel DC values extending above the point normally corresponding to full canopy. This feature appears in Fig. 1a, and is represented diagrammatically in Fig. 1b by the line segment between points c and d.
Points within the distribution between the bare soil line and the point corresponding to full canopy correspond to pixels with partial vegetation canopy (0 < GC < 1). Figure 1c shows an arbitrarily selected point e representing a pixel containing partial canopy. From previous discussions, we know that Eq. [3] and [4] define a curved line segment passing through point e that describes the transition of the pixel from bare soil (point f) to full canopy (point c). While the values of DCPixel,Red and DCPixel,NIR are known, and the values of DCFC,Red and DCFC,NIR can be determined by visually identifying the point in the distribution representing full canopy (point c), the values of DCSunlit-Soil,Red, DCShade-Soil,Red, DCSunlit-Soil,NIR, DCShade-Soil,NIR, and β are not apparent from the image data. Under these circumstances, it is not possible to explicitly solve Eq. [3] and [4] for the value of GC (i.e., to unmix the LMM).
In light of this difficulty, it is possible to suggest an approximate solution for the value of GC. As stated earlier, a simplification of the LMM represented by Eq. [3] and [4] is the LMM represented by Eq. [1] and [2]. For the example in Fig. 1c, this simplified LMM can be represented diagrammatically by the straight line between points f and c. Based on the available information, it is not possible to find the point along this straight line that directly corresponds to point e. However, an approximation would be a point on the straight line (point g) that is at the same perpendicular distance from the soil line as is point e. This perpendicular distance is indicated by the line segment between points g and h. The length of the curved line segment between points f and e, divided by the total length of the curved line between points f and c, is equal to the value of GC associated with point e. Thus, the length of the straight line segment between points f and g, divided by the total length of the straight line between points f and c, should be approximately equal to the value of GC associated with point e. Since the right triangle with vertices at points g, h, and f is similar to the right triangle with vertices at points c, i, and f, then the length of the line segment between points g and h, divided by the length of the line segment between points c and i, should also be approximately equal to the value of GC associated with point e. The length of the line segment between points g and h is equivalent to the value of the PVI calculated for point e,
 | [5] |
in which a1 and a0 are the slope and intercept, respectively, of the bare soil line (Richardson and Wiegand, 1977; Jackson et al.,1980, p.27; Liang, 2004, p. 255). Similarly, the length of the line segment between points c and i is equivalent to the value of PVI calculated for point c,
 | [6] |
where PVIFC refers to the value of PVI associated with full vegetation canopy. An approximate value for the GC associated with point e can now be calculated from the results of solving Eq. [5] and [6],
 | [7] |
From this theoretical analysis, we see that we only need the equation of the bare soil line and the red and NIR DC values associated with full canopy to estimate GC for pixels in a multispectral satellite image of an agricultural region. The bare soil line and the point associated with full canopy can usually be identified by plotting the DC values in the NIR spectral band vs. the corresponding DC values in the red spectral band for pixels comprising the image.
|
MATERIALS AND METHODS
|
|---|
A study was conducted during 2006 to assess the feasibility of using this approach to estimate GC of field crops using medium-resolution satellite imagery. Estimates of GC determined for a set of agricultural fields from Landsat-5 imagery were compared to corresponding ground-based observations of GC from the fields. Statistical analysis of these data was used to evaluate the accuracy of the satellite-based GC estimation approach.
Study Site
The study was conducted within a rectangular portion of the Texas High Plains containing 1000 km2 between the towns of Plainview and Floydada. The corners of the study area lay at 34°16' N, 101°42' W; 34°13' N, 101°20' W; 34°5' N, 101°45' W; and 33°57' N, 101°23' W. The region is predominantly farmland supporting field crops and pastures for cattle (Bos taurus). Based on hectares planted, the dominant field crop is cotton (Gossypium hirsutum L.), although corn (Zea mays L.) (for grain and silage), sorghum [Sorghum bicolor (L.) Moench] (for grain and forage), wheat (Triticum aestivum L.) (for grain and as a winter cover crop), alfalfa (Medicago sativa L.), sunflower (Helianthus annuus L.), and millet (Pennisetum glaucum L.) are also grown. The region is semiarid, with an average annual precipitation of 460 mm. Most of the field crops are irrigated, although some nonirrigated cotton is grown. Improved and unimproved pastures are found within the study area, and some fields are devoted to the irrigated production of grasses for forage, such as old world bluestem (Bothriochloa spp.), kleingrass (Panicum coloratum L.), and bermudagrass [Cynodon dactylon (L.) Pers.] or seed, such as side-oats grama [Bouteloua curtipendula (Michx.) Torr.]. Some land in the study area is in the Conservation Reserve Program (CRP) and supports a mixture of grasses and forbs. The predominant soils in the study area are nearly level to gently sloping noncalcareous clay loams and loams in the Pullman and Pullman-Olton associations (NRCS, 1974, 1978).
Within the study area, 31 agricultural fields were selected for evaluating the accuracy of the satellite-based GC estimation approach. These included 21 cotton fields (19 irrigated and two nonirrigated), five corn fields, two sorghum fields, and one alfalfa field. The number of fields in this sample devoted to each crop is roughly similar to the relative abundance of each field crop in the region, as described above. Periodically during the study, the fields were visited to obtain ground-based observations of GC using procedures described below. Field maps were marked to show the locations within each field where observations were made. In addition to the 29 fields planted to crops, two fields that remained bare over the growing season were included in the study to provide data for the situation of GC = 0.
Remote Sensing Data
Landsat-5 Thematic Mapper (TM) imagery containing the study site was acquired on four dates during the 2006 growing season: 30 June (Day 181), 16 July (Day 197), 1 August (Day 213), and 18 September (Day 261). Each image, located according to the Landsat World Reference System (WRS-2) along Path 30 at Row 36, was purchased on CD-ROM from the U.S. Geological Survey (USGS) EarthExplorer website (http://edcsns17.cr.usgs.gov/EarthExplorer/; verified 6 Nov. 2007). Pixel size in the imagery was specified as 30 m, and Systematic Correction (L1G) was applied by USGS to the image data. In Systematic Correction, the image is rotated, aligned, and georeferenced to a user-defined map projection (WGS84), and is radiometrically corrected based on characteristics of the TM sensor (Chander and Markham, 2003).
Each TM image was imported into ENVI image processing software (ITT, Boulder, CO). A 1140- by 975-pixel subset representing the study area was extracted from TM Band 2 (green spectral band), TM Band 3 (red spectral band), and TM Band 4 (NIR spectral band) of each image. The image subset was output as a standard 24-bit Windows bitmap (BMP) image file with the TM Band 4 data saved in the red channel of the BMP file, the TM Band 3 data saved in the green channel of the BMP file, and the TM Band 2 data saved in the blue channel of the BMP file. This resulted in a false-color rendition of the image that could be displayed in the manner of a standard color infrared photograph.
The BMP image for each acquisition date was imported into Photoshop (Adobe Systems, San Jose, CA). Each of the study fields was identified in the image, and the lasso tool was employed to delineate a subset of pixels that corresponded to the portion of the field where ground-based GC observations were made, as indicated by the field maps that were marked at the time of the field visits. The histogram function was then used to determine average DC values for the subset of pixels in the red, green, and blue channels of the BMP image. These corresponded to the average DC values in the NIR, red, and green spectral bands, respectively, of the original Landsat-5 image for that acquisition date.
In Photoshop, a 330- by 470-pixel subset was extracted from the center portion of each BMP image and was saved as a separate 24-bit BMP image. These subsetted BMP images were free of urban structures and contained an assortment of bare fields and fields with various amounts of vegetation, which made them ideal for determining the bare soil line and the value of PVIFC. The subsetted BMP images for 1 August and 18 September were cloud-free, but those for 30 June and 16 July contained isolated cumulus clouds. In Photoshop, the portions of the subsetted images containing clouds or cloud shadows were delineated using the lasso tool. With the paint bucket tool, pixel values in the delineated portions were then replaced with a DC of 255 in each of the three image channels, creating masked areas that appeared white in the displayed images. The Landsat-5 acquisition on 18 September occurred shortly after heavy rains in the region, so that numerous small lakes appeared scattered across the image. The procedure used to mask the subsetted BMP images for clouds was also used to mask the small lakes in the subsetted BMP image for 18 September. Each of the subsetted BMP images was then imported into ENVI, where the pixel DC values in each of the three image channels were output to an ASCII text file. The ASCII text files for the 30 June, 16 July, and 18 September acquisitions were then edited to remove data corresponding to the masked portions of the images.
Ground-based Observations
During the visits to each study field, ground-based observations of GC were obtained for comparison with corresponding values estimated using remote sensing. Only observations obtained from visits occurring within 3 d of a satellite image acquisition were used in this study. Three methods were available for determining ground-based observations of GC. Each method had advantages and disadvantages, which in part determined which method was used in a given field.
The first method was appropriate for contiguous rows of crop plants with closed leaf canopies, such as rows of mature cotton plants. In this case, the approximate width of the leaf canopy measured perpendicular to the row direction could be obtained with a meter stick. The average width of the leaf canopy could be determined from a set of individual width measurements (typically 20–30) made within the sampled portion of the field. The GC was determined by dividing the average leaf canopy width by the row spacing for the field. The main advantages of this method (called the "stick method") were that a large number of measurements could be quickly obtained, and that minimal processing of the data was required. The main disadvantage of this method was that it was not appropriate for open plant canopies, like corn or sorghum.
The second method used an AccuPAR Model PAR-80 Linear PAR Ceptometer (Decagon Devices, Pullman, WA) to make measurements of solar irradiance above and below the plant canopy. To determine GC, a measurement was initially made with the ceptometer held horizontally above the plant canopy to calibrate the linear array of 80 PAR sensors to the ambient irradiance level. The ceptometer was then held horizontally below the plant canopy at the soil level to allow the device to determine how many of the 80 PAR sensors were shaded. Dividing the number of shaded PAR sensors by 80 provided an estimate of GC. An average value for GC could be determined from a set of individual measurements (typically 20–30) made within the sampled portion of the field. The main advantages of this method (called the "ceptometer method") were similar to those of the stick method, except that it could be used for both open and closed plant canopies. The ceptometer method tended to overestimate GC when the sun was relatively low in the sky, so typically it was used during the period within a few hours of local solar noon. It could not be used on cloudy days.
The third method used a standard digital still camera mounted on a long pole to take overhead photographs of the field. The camera was mounted on the pole so that it could be positioned approximately 3 m above the ground pointed directly down at the plant canopy. Each digital image was imported into Photoshop. For fields with row crops (cotton, corn, and sorghum), the image was cropped to include only the row of plants directly below the camera. For alfalfa, the image was cropped to include only the central portion of the image. Cropping the images minimized the effects of optical distortions of the plant canopy present near the edges of the image. The lasso tool was used on the cropped image to delineate the portions containing leaf canopy. The histogram function was then used to determine the number of pixels in the delineated portions. Dividing the number of pixels in the delineated portions by the total number of pixels in the cropped image provided an estimate of GC. This method (called the "photo method") could be used for any crop, but the delineation of the leaf canopy in the images tended to be a tedious operation for open plant canopies, like corn or sorghum.
Data Analysis
The DC values in the NIR spectral band were plotted vs. the corresponding DC values in the red spectral band for pixel data from the edited ASCII text files for the 30 June, 16 July, 1 August, and 18 September Landsat-5 image acquisitions, resulting in distributions of points similar to that in Fig. 1a. A straight line was placed by visual inspection through the edge of each distribution at the location corresponding to bare soil, and the slope (a1) and intercept (a0) of each line was calculated. A point was identified visually at the top of each distribution at the location corresponding to full canopy, and the coordinates of these points (DCFC,Red, DCFC,NIR) were recorded.
For each acquisition date, the coordinates of the point corresponding to full canopy were used in Eq. [6] with the appropriate bare soil line slope and intercept values to calculate a PVI value corresponding to full canopy (PVIFC). Similarly, the average DC values in the red and NIR spectral bands determined for each of the 31 agricultural fields in the study (DCPixel,Red, DCPixel,NIR) were used to calculate a PVI value using Eq. [5]. Ground cover was then calculated for each field using Eq. [7]. Ground cover values were multiplied by 100 to obtain percent GC.
Ground cover values estimated for each field from the satellite image data were plotted vs. the corresponding values of GC obtained from ground-based field measurements. Simple linear regression (Ostle and Mensing, 1975, p. 169) was used to fit a straight line to these pairs of GC values. Student's t was then used to test if this regression line was significantly different from the 1:1 line (Ostle and Mensing, 1975, p. 177). Finally, a paired Student's t was used to test if the average GC value estimated from the satellite imagery was significantly different from the average GC value determined from ground-based field observations (Ostle and Mensing, 1975, p. 120). The average absolute error (AAE) was calculated between the values of GC estimated using the satellite image data (GCEST) and the corresponding GC values determined from the ground-based field observations (GCOBS) using the following formula,
 | [8] |
in which n is the number of observations (in this study, n = 51). The value of AAE indicates, on average, how close the estimated and observed GC values were, and provides an estimate of the overall accuracy of the procedure.
|
RESULTS
|
|---|
Results of plotting the DC values in the NIR spectral band vs. the corresponding DC values in the red spectral band for the 30 June, 16 July, 1 August, and 18 September Landsat-5 image acquisitions are presented in Fig. 2
. The distribution of points for each satellite acquisition date resembled the example in Fig. 1a, so the locations of the bare soil line and the point corresponding to full vegetation canopy could be identified for each distribution by visual inspection. The locations determined for these features are shown in Fig. 2. The slope (a1) and intercept (a0) of the bare soil line, along with the DC values in the red (DCFC,Red) and NIR (DCFC,NIR) spectral bands for full canopy, are summarized in Table 1
for the four satellite acquisitions. Also presented in this table are the PVI values associated with full canopy (PVIFC) calculated for each satellite acquisition using Eq. [6].
View larger version (41K):
[in this window]
[in a new window]
|
Fig. 2. Results of plotting the pixel digital count (DC) values in the near-infrared (NIR) spectral band vs. the corresponding DC values in the red spectral band for the 30 June, 16 July, 1 August, and 18 September Landsat-5 image acquisitions. The location of the bare soil line is indicated for each distribution of points, along with the location of the point representing full vegetation canopy ("FC").
|
|
View this table:
[in this window]
[in a new window]
|
Table 1. Slope (a1) and intercept (a0) of the bare soil line, digital count (DC) values in the red (DCFC,Red) and NIR (DCFC,NIR) spectral bands for full canopy, and the perpendicular vegetation index (PVI) value (PVIFC) associated with full canopy for the four Landsat-5 image acquisitions.
|
|
Results of estimating GC for the 31 agricultural fields in the study using the satellite image data are plotted in Fig. 3
vs. the corresponding ground-based field observations of GC. The points representing these data tend to lie along the 1:1 line. The least-squares linear regression line fit to these points has the following equation,
 | [9] |
where GCEST and GCOBS are expressed in percent. This regression explained 89.4% of the total variance among the points. Results of the Student's t test of the slope of this regression indicated that the slope was not significantly different from 1 (t = –1.546, 49 df, 
= 0.05). Similarly, the Student's t test of the intercept of this regression indicated that the intercept is not significantly different from 0 (t = 0.107, 49 df, = 0.05). Thus, it is suggested that the regression line through these points is not significantly different from the 1:1 line.
View larger version (27K):
[in this window]
[in a new window]
|
Fig. 3. Values of ground cover (GC) estimated for the 31 agricultural fields in the study using the satellite image data plotted vs. the corresponding ground-based field observations of GC. The solid diagonal line represents the 1:1 line, while the dashed line represents the least-squares linear regression line fit to these points.
|
|
The average value of GCEST was 43.95% while the average value of GCOBS was 43.82%. The Student's t test of the pairs of estimated and observed GC values indicates that the average value of GCEST is not significantly different from the average value of GCOBS (t = –0.110, 50 df, = 0.05). The calculated value of AAE for this data set was 5.76% This suggests that, on average, estimates of GC determined using this procedure were within 6% of their true values.
|
DISCUSSION
|
|---|
As these results indicate, the procedure was able to accurately estimate the ground cover of various agricultural crops from medium-resolution satellite image data. As might be expected, the accuracy of a set of GC estimates derived from a given satellite image depends on the identification of the soil line and point corresponding to full canopy in the distribution of red and NIR pixel DC values. Identification of these features was easy for the image data from the four acquisition dates in this study (Fig. 2), since the shapes of the distributions of points closely resembled the theoretical model outlined in Fig. 1b. This close match was achieved because the portions of the Landsat-5 images used to develop the distributions were selected to contain almost exclusively agricultural targets (vegetation, bare soil, and a mixture thereof). The inclusion of nonagricultural targets (such as buildings, paved surfaces, water bodies, clouds, and cloud shadows) in this analysis can confound the identification of the bare soil line and the point corresponding to full canopy. For example, Fig. 4a
shows the distribution of points representing DC values in the red and NIR spectral bands for pixels in the portion of the 30 June Landsat-5 image used to construct the distribution for that date shown in Fig. 2, except that in this case the pixels containing clouds and/or cloud shadows were not masked out before constructing the distribution. In Fig. 4a, it is much more difficult to identify where the bare soil line should be located, due to the points corresponding to pixels affected by clouds and cloud shadows. Similarly, Fig. 4b shows the distribution of points corresponding to the 18 September Landsat-5 image shown in Fig. 2, except that in this case the pixels containing lakes were not masked out before constructing the distribution. Again, the idealized form of the distribution expected from Fig. 1b is confounded by the points corresponding to pixels affected by the lakes. Therefore, proper screening of the medium-resolution multispectral satellite imagery is important to remove nonagricultural targets and facilitate identification of the features necessary for applying this procedure.
View larger version (26K):
[in this window]
[in a new window]
|
Fig. 4. Results of plotting the pixel digital count (DC) values in the near-infrared (NIR) spectral band vs. the corresponding DC values in the red spectral band for the (A) 30 June and (B) 18 September Landsat-5 image acquisitions without masking the image data to remove nonagricultural targets such as clouds, cloud shadows, and lakes.
|
|
Since there are almost always some bare soil surfaces present in most agricultural regions at any time, it is usually not difficult to identify the bare soil line in the distribution of points when the nonagricultural targets have been removed. In the region where this study was conducted (the Texas High Plains), it is possible to find vegetation of a sufficient density in a satellite image to allow identification of the point corresponding to full canopy at practically any time during the growing season. Even in the early spring, winter wheat canopies are sufficiently dense to allow this determination. However, it is conceivable that in some agricultural regions there might be periods at the start or end of the growing season for which there is no vegetation with a density approaching full canopy, thereby making identification of the full canopy point difficult. In such cases, one could use an average value of PVIFC determined at other times during the growing season in calculating GC. For example, the average of the values of PVIFC presented in Table 1 is 96.1. Using this average value in the calculations of GC for each of the four acquisition dates in this study results in an average GC of 48.06% which is <5% different from the corresponding value (GCEST = 43.95%) obtained using the values of PVIFC from Table 1 that are specific to each acquisition date. Therefore, in the absence of adequate data to identify the point corresponding to full canopy, it may be feasible to use a previously determined average value of PVIFC in the calculations of GC.
Application of this procedure currently requires that the bare soil line and the point corresponding to full vegetation canopy be specified by visual inspection of the distribution of NIR and red DC values. While this is generally not difficult when the nonagricultural targets have been masked from the imagery, it does involve a degree of subjectivity. It also hinders automation of the procedure. The development of mathematical procedures for masking nonagricultural targets and for objectively identifying the bare soil line and the point corresponding to full vegetation canopy could allow complete automation of this approach.
|
CONCLUSIONS
|
|---|
The results of this study demonstrate that GC of field crops can be accurately estimated using medium-resolution multispectral satellite imagery without relying on empirical relationships or calibration of the satellite image data. Plotting pixel DC values in the NIR spectral band vs. corresponding DC values in the red spectral band allows determination of the equation of the bare soil line, which allows calculation of the PVI for any pixel in the image. It also allows determination of the point corresponding to full vegetation canopy, which allows calculation of the PVI value corresponding to full canopy. The GC for the pixel is then estimated as the ratio of these two PVI values. Removing pixels corresponding to nonagricultural targets (such as buildings, paved surfaces, water bodies, clouds, and cloud shadows) from the satellite image data simplifies the application of this procedure. Comparing estimated and observed values of GC suggests that, on average, estimates of GC determined using this procedure should be within 6% of their true values.
|
ACKNOWLEDGMENTS
|
|---|
The authors wish to recognize the invaluable support of this research by the Texas Alliance for Water Conservation (TAWC) Demonstration Project funded by the Texas Water Development Board (TWDB).
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.
|
REFERENCES
|
|---|
- Asrar, G., R.B. Myneni, and E.T. Kanemasu. 1989. Estimation of plant canopy attributes from spectral reflectance measurement. p. 252–296. In G. Asrar (ed.) Theory and applications of optical remote sensing. John Wiley & Sons, New York.
- Chander, G., and B. Markham. 2003. Revised Landsat-5 TM radiometric calibration procedures and postcalibration dynamic ranges. IEEE Trans. Geosci. Remote Sens.
41
(11):2674–2677.
- Gausman, H.W. 1985. Plant leaf optical properties in visible and near-infrared light. Texas Tech Press, Lubbock.
- Hatfield, J.L., J.H. Prueger, and W.P. Kustas. 2004. Remote sensing of dryland crops. p. 531–568. In S.L. Ustin (ed.) Remote sensing for natural resource management and environmental monitoring: Manual of remote sensing. 3rd ed. Vol. 4. John Wiley & Sons, New York.
- Heilman, J.L., W.E. Heilman, and D.G. Moore. 1982. Evaluating the crop coefficient using spectral reflectance. Agron. J.
74
:967–971.[Abstract/Free Full Text]
- Jackson, R.D., P.J. Pinter, Jr., R.J. Reginato, and S.B. Idso. 1980. Hand-held radiometery. Agricultural reviews and manuals ARM-W-19. USDA-ARS, Oakland, CA.
- Jensen, J.R. 2000. Remote sensing of the environment, an earth resource perspective. Prentice Hall, Upper Saddle River, NJ.
- Liang, S. 2004. Estimation of land surface biophysical variables. p. 246–309. In Quantitative remote sensing of land surfaces. John Wiley & Sons, New York.
- Maas, S.J. 1997. Structure and reflectance of irrigated cotton leaf canopies. Agron. J.
89
:54–59.[Abstract/Free Full Text]
- Maas, S.J. 1998. Estimating cotton canopy ground cover from remotely sensed scene reflectance. Agron. J.
90
:384–388.[Abstract/Free Full Text]
- Maas, S.J. 2000. Linear mixture modeling approach for estimating cotton canopy ground cover using satellite multispectral imagery. Remote Sens. Environ.
72
:304–308.
- Maas, S.J., N. Rajan, J. Ko, J. Duesterhaus, and R. Lascano. 2005. Remote sensing approach for estimating daily crop water use. In C. Yang and J.H. Everitt (ed.) Proc., 20th Biennial Workshop on Aerial Photography, Videography, and High Resolution Digital Imagery for Resource Assessment, Weslaco, TX. 4–6 Oct. 2005. Am. Soc. Photogramm. Remote Sens., Bethesda, MD.
- Mahiny, A.S., and B.J. Turner. 2007. A comparison of four common atmospheric correction methods. Photogramm. Eng. Remote Sens.
73
(4):361–368.
- Natural Resource Conservation Service. 1974. Soil survey of Hale County, Texas. NRCS, USDA, Washington, DC.
- Natural Resource Conservation Service. 1978. Soil survey of Floyd County, Texas. NRCS, USDA, Washington, DC.
- Ostle, B., and R.W. Mensing. 1975. Statistics in research. Iowa State Univ. Press, Ames, IA.
- Pinter, P.J., Jr., J.L. Hatfield, J.S. Schepers, E.M. Barnes, M.S. Moran, C.S.T. Daughtry, and D.R. Upchurch. 2003. Remote sensing for crop management. Photogramm. Eng. Remote Sens.
69
(6):647–664.
- Rajan, N., and S.J. Maas. 2006. Estimating daily and seasonal crop water use of High Plains cropping systems using remote sensing and crop modeling. p. 25–29. In R.C. Schwartz et al. (ed.) Proc., 28th An. Southern Conserv. Systems Conf., Amarillo, TX. 26–28 June 2006. USDA-ARS, Bushland, TX.
- Richardson, A.J., and C.L. Wiegand. 1977. Distinguishing vegetation from soil background information. Photogramm. Eng. Remote Sens.
43
(12):1541–1552.
- Richardson, A.J., C.L. Wiegand, D.F. Wanjura, D. Dusek, and J.L. Steiner. 1992. Multisite analysis of spectral-biophysical data for sorghum. Remote Sens. Environ.
41
:71–82.
- Wiegand, C.L., A.H. Gerbermann, K.P. Gallo, B.L. Blad, and D. Dusek. 1990. Multisite analyses of spectral-biophysical data for corn. Remote Sens. Environ.
33
:1–16.
- Wiegand, C.L., S.J. Maas, J.K. Aase, J.H. Hatfield, P.J. Pinter, Jr., R.D. Jackson, E.T. Kanemasu, and R.L. Lapitan. 1992. Multisite analyses of spectral-biophysical data for wheat. Remote Sens. Environ.
42
:1–21.
TOP
ABSTRACT
INTRODUCTION
