| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
a Laboratorio Nacional de Modelaje y Sensores Remotos, INIFAP, km 32.5 Carr. Aguascalientes-Zacatecas, Ap. Postal 20 Pabellon de Arteaga, Aguascalientes 20660, Mexico
b USDA-ARS, Grassl. Soil and Water Res. Lab., 808 East Blackland Rd., Temple, TX 76502, USA
c Dep. of Plant and Soil Sci., Texas Tech Univ., 3810 4th St., Lubbock, TX 79415, USA
d Campo Experimental de Valle del Fuerte, INIFAP, Carr. Internal. Mexico-Nogales, km 609 Ejido San Jose Rios, Guasave Sinaloa 81200, Mexico
e Campo Experimental de Rio Bravo, INIFAP, km 61 Carr. Matamoros-Reynosa, Ap. Postal 172, Rio Bravo, Tamaulipas 88900, Mexico
f Campo Experimental de Culiacan, INIFAP, Carr. Culiacan El Dorado, km 17.5 Culiacan, Sinaloa 80000, Mexico
* Corresponding author (abaez{at}labpred.inifap.gob.mx)
Received for publication December 17, 2003.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Abbreviations: DAS, days after sowing • ETM, enhanced thematic mapper • GCOS, Global Climate Observation System • gLAI, ground-measured leaf area index • GTOS, Global Terrestrial Observation System • LAI, leaf area index • LCS, lack of correlation • MSD, mean squared deviation • NDVI, normalized difference vegetation index • NIR, near infrared • R, red (wavelength band) • RMSE, root mean square error • SAVI, soil-adjusted vegetation index • SB, squared bias • SDSD, squared difference between standard deviations • sLAI, satellite-derived leaf area index
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Crop models have been used for monitoring crop growth and predicting yield (Baez-Gonzalez and Jones, 1995a, 1995b; Sinclair and Seligman, 1996; Sivakumar and Glinni, 2002). However, their use in large areas has been limited because the required inputs are generally available only at field scale. Remote sensing provides observations over large areas at regular intervals, making it useful in large-scale crop modeling (Gallo and Flesch, 1989; Moulin et al., 1998; Reynolds et al., 2000). Numerous studies have been conducted on its use in assessing crop growth and yield at regional and national levels (Hamar et al., 1996; Denore et al., 2000; Bochenek, 2000).
Remote sensing in crop modeling refers to quantification of a plant community attribute obtained with instruments that are not in contact with the plants. This often involves the measurement of electromagnetic radiation in specific wavelengths reflected or emitted by the plants (Maas, 1988). Remote sensing has been used to estimate crop parameters such as photosynthetic rate, intercepted photosynthetically active radiation (IPAR), biomass, photosynthetic size of canopy, and LAI (Ochi et al., 2000; Aparicio et al., 2002; Lobell et al., 2002).
Leaf area index, a quantitative measure of foliage density, is a key variable in agricultural modeling. It was originally defined as the ratio of leaf area to a given land area (Hay and Walker, 1989). A more recent definition covers vegetation with different photosynthetic or morphological characteristics: LAI is half the all-sided living foliage per unit ground surface area projected on the horizontal datum (Fernandes et al., 2003). Through satellite remote sensing, data on the distribution of LAI over wide areas may be obtained (Wiegand et al., 1979; Wiegand and Richardson, 1984; Chen and Cihlar, 1996). A combination of near-infrared (NIR) and red (R) reflectance defined as normalized difference vegetation index (NDVI) has been used to indirectly estimate LAI (Asrar et al., 1984). Use of NDVI leads in remote sensing applications despite the emergence of more theoretically reliable indices such as the soil-adjusted vegetation index (SAVI), transformed SAVI (TSAVI), and the atmospherically resistant vegetation index (MSAVI) (Rondeaux et al., 1996; Haboudane et al., 2002). Many studies have been made on the relationship between LAI and NDVI (Gower et al., 1999; Turner et al., 1999; Qi et al., 2000). While there have been studies on LAI in relation to crop monitoring and assessment (Pollock and Kanemasu, 1979; Colombo et al., 2003), there is a need to explore further its use in large-area yield prediction.
In Mexico, a national yield prediction project has been established by the National Research Institute of Forestry, Agriculture, and Animal Production (INIFAP) with logistical support from the USDA-ARS and Texas Agricultural Experiment Station (TAES) for the purpose of predicting yield for maize and other crops at regional scale. Several approaches have been developed for yield prediction, one of which involves the use of a yield model based on LAI. Since ground measurements of LAI are difficult to obtain for large-area yield prediction, our goal was to explore the use of sLAI.
The objectives of this study were (i) to derive LAI from high spatial resolution imagery data with the aim of using such data to replace ground measurements and (ii) to forecast maize yield in the middle of the growing season in large areas (<100000 ha) using a simple regression yield model based on LAI. This study focuses on large-area prediction in two important agricultural regions of Mexico.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
The study focuses on Sinaloa's northern and central regions (Fig. 1) , which have different agroclimatic patterns. The northern region (26°21' N, 109°16' W to 25°22' N, 108°16' W; mean elevation of 71 m) has a mean annual temperature of 25.4°C (min. 5°C, max. 43.5°C) and a mean annual precipitation of 421.8 mm while the central region (24°58' N, 107°54' W to 24°02' N, 107°03' W; mean elevation of 32 m) has a mean annual temperature of 23.8°C (min. –2.0°C, max. 41.7°C) and a mean annual precipitation of 707.9 mm (SAGARPA, 2003).
|
 | [1] |
Ground-measured LAI and yield data of three consecutive autumn–winter maize growing seasons (1999–2000, 2000–2001, and 2001–2002) were used to validate the model. Measurements of LAI and yield were collected from sites (farmers' fields) located in the northern and central valleys of Sinaloa. Fifteen sites (seven north, eight central) were used in 1999–2000, seven in 2000–2001 (all in the central region), and 11 (nine north, two central) in 2001–2002. The selected sites ranged from 8 to 300 ha and had 90% or greater homogenous maize cover. All sites were irrigated. A sampling point was permanently established at each site, and the global positioning system (GPS) was used to locate the sites. A Decagon AccuPAR (Decagon Devices, Inc., Pullman, WA) was used to sample LAI every 2 wk from the eight-leaf stage up to the silking stage. Wilhelm et al. (2000) found a high correlation between destructive sample LAI values and AccuPAR estimates for maize; in their results for two hybrids, correlation coefficient values were 0.93 and 0.86 for AccuPAR LAI estimates as a function of destructive LAI estimates. Grain yield (Mg ha–1) was measured in each field by destructive methods on the same day that the farmers harvested the fields. At each site, two rows (5 m each) of maize were harvested, and a sample of grain yield was collected. Grain weight was reported at 140 g kg–1 of moisture. The validation of the yield model is further discussed in Results.
Satellite-Derived Leaf Area Index
To use Eq. [1] as a large-scale yield predictor, an estimate of maximum LAI for large areas is needed. For this purpose, an empirical regression model was derived from another data set of 56 gLAI values (41 and 15 for central and northern regions, respectively) obtained during the 2001–2002 autumn–winter growing season, following the same procedures described earlier. These values were correlated with the same number of NDVI values derived from the NIR and R wavelength bands of Landsat-7 ETM+ for the corresponding field locations. The images covered the central (Row 42, Path 33 and Row 43, Path 33) and northern (Row 43, Path 32) regions of Sinaloa. Images of 8 Feb. 2002 [84 ± 10 mean d after sowing (DAS)] and 5 Mar. 2002 (103 ± 6 mean DAS) were selected for the central and northern regions, respectively, because they were closest to the time that maize in Sinaloa usually reaches maximum LAI (100 to 115 DAS in the central region and 115 to 125 DAS in the northern region) (Baez-Gonzalez et al., unpublished data, 2003).
Image digital count data used in calculating NDVI were Level 1 Systematic Corrected by the data provider. No additional corrections for earth–sun distance, solar zenith angle, or atmospheric clarity were applied to the digital count data. The NDVI was calculated as (NIR – R)/(NIR + R). The sLAI model that was developed is presented in Results.
Validation of Yield and Satellite-Derived Leaf Area Index Models with 2002–2003 Data
The yield model was further validated with both gLAI and sLAI data of 2002–2003. Ground data were gathered during the 2002–2003 autumn–winter growing season from 71 farmers' fields in Sinaloa (36 fields in the northern part and 35 in the central area, Fig. 1). The selected fields represented the crop conditions for approximately 300000 ha of maize in the agricultural valleys of Sinaloa.
The latitude and longitude of each field were recorded. This enabled us to collect LAI at the same location for repeated sampling and to select the appropriate pixel from the satellite image. Planting dates ranged from 31 October to 14 December in the northern part and from 28 October to 15 December in the central part. Starting in January 2003, LAI was measured every 15 d until the crop reached silking stage.
Grain yield was predicted using Eq. [1] first with maximum gLAI and next with sLAI, which was derived from Landsat-7 ETM+ images using Eq. [2]. The two images used for the central region (Row 42, Path 33 and Row 43, Path 33) were of 19 Jan. 2003 (67 ± 10 mean DAS); the image for the north (Row 43, Path 32) was of 27 Feb. 2003 (98 ± 12 mean DAS). Later images would have been ideal; however, the February 2003 images that we obtained for the central region could not be used because of clouds covering most of the area. Since the peak growth period (100–115 DAS in the central region, 115–125 DAS in the north) is so essential, we have to ensure that in future applications, imagery from some satellite is acquired around that period. Seventy-one NDVI values (36 for the north, 35 for the central part) were used to obtain the sLAI.
Data Analysis
To validate the yield model, the measured and simulated grain yields were regressed, and the root mean square error (RMSE) computed as described by Ahuja and Ma (2002).
We used SAS GLM (SAS Inst., 1985) procedures for the regression analyses of sLAI. We analyzed how closely the sLAI matched the gLAI gathered closest to the dates of the satellite images and how well sLAI matched maximum gLAI. The deviation (measured minus derived) for sLAI was examined to see how many of the values exceeded the predetermined criteria (Mitchell, 1997; Mitchell and Sheehy, 1997) of ±1.0 for threshold performance and ±0.5 for optimum performance. These criteria are within the range established by the Global Climate Observation System (GCOS) and the Global Terrestrial Observation System (GTOS) (Fang et al., 2003). The gLAI gathered in the period of 10 to 25 Feb. 2003 and on 27 Jan. 2003 were used for the northern and central regions, respectively. The RMSE values were calculated as described by Ahuja and Ma (2002).
For comparisons of the model performance using maximum gLAI and sLAI, the mean squared deviation (MSD) for grain yield was calculated as well as its three components: the squared bias (SB), the lack of correlation weighted by the standard deviation (LCS), and the squared difference between standard deviations (SDSD) (Kobayashi and Salam, 2000). An analysis was made of the following: predicted yield using maximum gLAI and sLAI vs. actual yield and the accuracy of model predictions using maximum gLAI and sLAI in the northern and central regions.
|
RESULTS AND DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
To derive remotely sensed LAI, the following empirically derived logarithmic regression model was developed (Fig. 2)
:
 | [2] |
|
= 0.05). The slope of the equation was not significantly different from 1.0, and the 
intercepted was not significantly different from 0.0. The NDVI was used for this study because it correlates well with foliage density (Rondeaux et al., 1996) and has been demonstrated to give satisfactory LAI estimates (Qi et al., 2000). The approach of determining LAI by establishing a relationship between NDVI and LAI is widely used because of its simplicity and ease of computation. (Colombo et al., 2003; Qi et al., 2000). However, the approach has its limitations, one of which is that the sensitivity of NDVI to LAI becomes weaker with increasing LAI values (Baret and Guyot, 1991). Another is the sensitivity of NDVI to nonvegetation-related factors such as solar and viewing geometry, soil background, and atmospheric conditions (Baret and Guyot, 1991; Rondeaux et al., 1996). Previous studies have likewise shown that the relationship between LAI and spectral vegetation indices is affected by such factors as background reflectance and stand age (Colombo et al., 2003; Moreau et al., 2003; Butson and Fernandes, 2004).
For this study, calculations indicated that approximately a 10% difference in radiance as measured at the top of the atmosphere existed between the two Landsat overpass dates (8 February and 5 March). This difference is the result of differences in earth–sun distance and solar zenith angle between these two dates at these latitudes. Use of NDVI values in Eq. [2] effectively compensates for this difference. In this study, additional information required to correct the satellite data for atmospheric clarity was not available. Therefore, it is possible that a portion of the scatter in points around the regression line in Fig. 2 is the result of differences in atmospheric clarity between sites and dates.
Satellite-Derived Leaf Area Index vs. Ground-Measured Leaf Area Index
Results involving the use of Eq. [2] with the 2002–2003 validation data set showed a satisfactory match between sLAI and gLAI. The overall RMSE (0.90) is within the acceptable range established by GCOS and CTOS. In Fig. 3A
, we can see that Eq. [2] met the threshold criterion (±1.0) in 25 of 36 sites (75%) in the north and in 22 of 35 sites (62%) in the central region. The mean and standard deviation in the northern region was 4.93 ± 0.63 for gLAI and 4.92 ± 0.89 for sLAI. In the central region, the mean and standard deviation was 4.90 ± 0.47 for gLAI and 4.70 ± 0.97 for sLAI. The RMSE of the northern and central regions was 0.87 and 0.92, respectively. On the whole, these results indicate that Eq. [2] can adequately estimate gLAI on the date of the satellite image.
|
The deviation analysis of sLAI in relation to maximum gLAI at the 71 sites showed that in 62% of the cases, the value of sLAI was within the range of 1.0 of the value of maximum gLAI (Fig. 3B). The negative values in Fig. 3B signify cases when maximum gLAI was higher than sLAI while positive values show the opposite. Only the northern region showed sLAI values higher (<+1.0) than the maximum gLAI values. In the central region, 13 of the 35 sites showed negative values of <–1.0, which contrasts with the number in the northern region (7 out of 36). These results were expected since the 2003 Landsat images (the available ones that were noncloudy and closest to the peak period) captured areas in the central region when the crop was still in the early stage of development. The selection of images is thus crucial; for greater accuracy, the dates of images must match as closely as possible the dates when the crop reaches maximum LAI.
The RMSE was 1.0 and 1.1 in the northern and central regions, respectively. On the whole, results seem to indicate that Eq. [2] can be used to derive LAI from high spatial resolution images to avoid costly field data in Sinaloa, but some degree of error (RMSE = 1.1) in estimation of LAI values can be expected.
Yield Prediction and Model Performance
This section presents results in relation to the second objective of the study, which was to predict yield using the LAI-based yield model.
Measured and simulated grain yields were compared to validate the yield model (Table 1). Validation is taken here to mean checking if the model's outputs are sufficiently close to the observed data and if the model works with totally independent data sets; that is, it accurately predicts yield (Boote et al., 1996). Validation is an essential process for models that are applied, with predictions used to replace costly field measurements (Mitchell, 1997). The following acceptable simulation error levels at different stages of the growing season have been predetermined based on our experience with yield forecasting in Mexico: 30% during the preplanting period (using weather forecast data), 20% midseason (using LAI-based yield model), and 10% one month before harvest (using crop growth model with satellite data). This study focused on prediction in the middle of the growing season (i.e., when the crop reaches maximum LAI) in large areas (over 100000 ha) in Sinaloa.
|
The results of the second validation using the 2002–2003 data set are likewise presented in Table 1. This time the model was validated with both gLAI and sLAI. With maximum gLAI, the model was able to predict with a good level of accuracy in the whole study area (mean simulation error of –9.3%). The accuracy was higher in the central region (mean simulation error of –7.3%) than in the northern part (mean simulation error of –11.4%). One possible reason is the variability of grain yield in the regions. In the northern region, the measured yield showed a higher variability, with a standard deviation of ±2.2 Mg ha–1 compared with ±1.6 Mg ha–1 in the central region. To illustrate this variability, in the northern region, the reported maximum gLAI of Sites 24, 27, and 32 was 4.9, 5.0, and 5.0, respectively (data not shown). These sites showed almost the same LAI; however, their measured grain yield varied: 12.2 Mg ha–1 in Site 24, 6.8 Mg ha–1 in Site 27, and 10.0 Mg ha–1 in Site 32. On the other hand, in the central region, Sites 23 and 29 had almost the same maximum gLAI values (7.0 for Site 23 and 6.9 for Site 29) and measured yields (11.0 and 10.7 Mg ha–1, respectively). This indicates that green biomass is not the sole predictor of yield and that there are other factors involved such as growing conditions.
The replacement of maximum gLAI by sLAI increased the mean simulation error of the yield model by 2% (from –9.3 to –11.1%). Its effect differed in the two regions. In the north, the replacement did not significantly affect the simulation error (from –11.4 to –10.4%). On the other hand, in the central region, the mean simulation error increased by nearly 5% (from –7.3 to –11.9%) when sLAI was used as model input. This may be partly attributed to the early satellite image used in this area. The same sites mentioned earlier illustrate how the replacement of gLAI by sLAI affected the accuracy of the yield model. In the northern region, the simulation error for Sites 24, 27, and 32 was –26.0, 32.9, and –12.1%, respectively, with maximum gLAI and –25.4, 27.6, and –19.7%, respectively, with sLAI, indicating that the replacement of maximum gLAI by sLAI had only a minor effect on model accuracy. But in the central region, the simulation error for Sites 23 and 29 was –7.2 and –4.4%, respectively, with maximum gLAI and –18.6 and –13.8% with sLAI.
On the whole, when sLAI was used instead of gLAI, the increase in mean absolute error in terms of grain yield was 0.3 Mg ha–1 (–1.5 Mg ha–1 with sLAI vs. –1.2 Mg ha–1 with gLAI). With this minimal increase in error, the replacement of gLAI by sLAI is considered feasible.
To measure the overall deviation of the yield model, the MSD for the full study area was calculated. As expected, the MSD value was lower when the model used maximum gLAI (4.4 vs. 5.8). The SB values (1.6 for max. gLAI, 2.3 for sLAI) indicate that the change of the source of LAI affects the accuracy of the model. The mean and standard deviation of predicted yield was 9.3 ± 0.7 Mg ha–1 with maximum gLAI and 9.0 ± 0.4 Mg ha–1 with sLAI. The measured yield in the whole study area was 10.5 ± 1.9 Mg ha–1.
The use of two different types of LAI input resulted in similar values for LCS (1.35 vs. 1.34) weighted by the standard deviation and for SDSD. This indicates that the model simulated in a similar way the pattern of fluctuations across measurements in the whole study area regardless of the type of input. However, the magnitude of yield fluctuations in the full study area was better simulated when maximum gLAI was used (SDSD of 1.5 for maximum gLAI and 2.2 for sLAI).
For the northern region, the yield model had lower overall deviation (MSD 5.2 vs. MSD 6.5, Table 1) when it used maximum gLAI. The SB showed the same values (2.1 vs. 2.1) for maximum gLAI and sLAI. This shows that in the northern sites, the change in the source of LAI did not affect the accuracy of the yield model. The mean and standard deviation of simulated grain yield was 9.0 ± 0.8 Mg ha–1 with maximum gLAI and 9.0 ± 0.4 Mg ha–1 with sLAI. The mean and standard deviation of measured yield in the northern region was 10.5 ± 2.2 Mg ha–1. In terms of LCS values, the yield model had almost the same values for the two types of LAI (1.2 and 1.3 for maximum gLAI and sLAI, respectively), indicating that the model simulated the pattern of fluctuations across measurements in the same way regardless of the type of LAI. However, based on the SDSD value, the yield model simulated the magnitude of fluctuations better when it used maximum gLAI (1.9 against 3.0).
For the central region, the yield model also had lower overall deviation when it used maximum gLAI (maximum gLAI MSD = 3.3 vs. sLAI MSD = 4.9) (Table 1). In the case of the MSD of the yield model using sLAI, the SB was the component that contributed the most to its value. The SB values (0.9 with maximum gLAI and 2.2 with sLAI) indicate that in the central region, the accuracy of the yield model was affected by the source of LAI. The mean and standard deviation of simulated grain yield was 9.5 ± 0.4 Mg ha–1 with maximum gLAI and 9.0 ± 0.4 Mg ha–1 with sLAI. The mean and standard deviation of measured yield was 10.5 ± 1.6 Mg ha–1. In this region, the yield model simulated in a similar way the pattern and magnitude of fluctuations across measurements whether using maximum gLAI or sLAI. This is shown by the LCS values (1.1 for maximum gLAI and 1.3 for sLAI) and the SDSD values (1.3 and 1.4 for maximum gLAI and sLAI, respectively) in Table 1.
|
CONCLUSIONS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
Overall results show that the use of sLAI in place of measured LAI increased the mean simulation error of the yield model by 2% (from –9 to –11%). In the northern region, the error was almost the same with either source of LAI (from –11 to –10%) while in the central region, the mean simulation error increased by nearly 5% (from –7 to –12%) when sLAI was used as model input, possibly because of the early image used. Further testing needs to be done to determine the consistency of performance of the sLAI model.
The low temporal resolution of some types of satellite imagery such as Landsat-7 ETM+ has long been considered a limitation in the use of remote sensing for agricultural purposes (Inoue, 2003). For more accurate yield forecasting using the approach developed in this study, satellite images should be close in time to the occurrence of the maximum leaf area. It is essential to get satellite imagery (Landsat, SPOT, IRS, etc.) at the peak growth period. Otherwise, a model to predict LAI has to be devised that could be calibrated at various times during the growing season using image data. Also, for better sLAI estimation, the use of other vegetation indices such as the optimized SAVI needs to be explored.
Finally, while the study results indicate that the use of sLAI in place of gLAI with the yield model will result in an increase in simulation error, its use can still be feasible considering that it will eliminate laborious and costly ground estimation for midseason large-area yield prediction in Sinaloa.
|
ACKNOWLEDGMENTS |
|---|
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSREFERENCES |
|---|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 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 | |||
