- Email: [email protected]
Spatial validation of crop models for precision agriculture
Agricultural Systems 68 (2001) 97±112
www.elsevier.com/locate/agsy
Spatial validation of crop models for precision agriculture B. Basso a,*, J.T. Ritchie b, F.J. Pierce c, R.P. Braga d, J.W. Jones d a Dipartimento di Produzione Vegetale, Universita' degli Studi della Basilicata, 85100 Potenza, Italy Department of Crop and Soil Sciences, Michigan State University, East Lansing, MI 48824-1325, USA c Center for Precision Agricultural Systems, Washington State University Prosser, WA 99350, USA d Department of Agricultural Engineering, University of Florida, Gainsville, FL 32611, USA
b
Received 31 July 2000; received in revised form 11 October 2000; accepted 24 October 2000
Abstract Spatial measurements of yield using technological advances like on-the-go yield monitoring systems have clearly shown large within-®eld variability in crop yields suggesting that ®eld yields could be increased or cost decreased by varying management over space. This study evaluated the utility of the CROPGRO-Soybean simulation model and remote sensing in the interpretation of a soybean yield map. CROPGRO was executed on areas within the ®eld de®ned as reasonably uniform by a Normalized Dierence Vegetative Index (NDVI) analysis. The model was able to closely predict the crop yield variability measured within the ®eld when the measured soil type and plant population were used as model inputs. Remote sensing was useful in ®nding spatial patterns across the ®eld to target sampling and to provide spatial inputs for the model. Results of this study showed that a combination of crop model and remote sensing can identify management zones and causes for yield variability, which are prerequisites for zone-speci®c management prescriptions. # 2001 Elsevier Science Ltd. All rights reserved. Keywords: Crop yield spatial variability; Crop models; Remote sensing; NDVI; Management zones
* Corresponding author. Tel.: +39-971-202-271; fax: +39-971-202-269. E-mail address: [email protected] (B. Basso). 0308-521X/01/$ - see front matter # 2001 Elsevier Science Ltd. All rights reserved. PII: S0308-521X(00)00063-9
98
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
1. Introduction Agricultural production systems are inherently variable due to spatial variation in soil properties, topography, climate and other factors. To achieve the ultimate goal of sustainable cropping systems, variability must be considered both in space and time because the factors in¯uencing crop yield have dierent spatial and temporal behavior (Verghagen and Bouma, 1997; Pierce and Nowak, 1999). Advances in technologies such as Global Positioning Systems (GPS), Geographic Information Systems (GIS) and remote sensing have created the possibility to assess the spatial variability present in the ®eld and manage it with appropriate site-speci®c practices (Pierce et al., 1997; Verhagen et al., 1995). Site-speci®c management (SSM) strategies may be able to optimize production, but their potential bene®ts are highly dependent on the accuracy of the assessment of such variability (Pierce and Nowak, 1999). Traditional analytical techniques, such as regression of static measurements against yield, have failed to explain the reasons for yield variability because the dynamic, thus temporal, interactions of stresses on crop growth and development cannot be accounted for (Jones and Ritchie, 1991; Cambardella et al., 1996; Sudduth et al., 1996). Process oriented crop simulation models, such as the CERES and CROPGRO (Ritchie et al., 1985; Boote et al., 1998), integrate the eects of temporal and multiple stress interactions on crop growth processes under dierent environmental and management conditions. Even though crop models have shown high potential for optimizing production and minimizing environmental impact (Ritchie, 1987, 1991), their application for SSM has been limited thus far (Sadler et al., 2000). Crop models can be used for understanding yield variability in both space and time, leading to a more sustainable environment (Sadler and Russell, 1997; Cora et al., 1999). Batchelor et al. (1998) and Paz et al. (1997, 1999) used CERES-Maize and CROPGRO-Soybean simulation models to determine the eect of soil moisture variation throughout the season on yield spatial variability optimizing for soil water limit parameters. The differences between measured and predicted yield for 224 grid points over a 3-year period ranged from 10% for 70% of the grids and 20% for 96% of the grids in maize and for soybean from 10% for 84% of the grids and 20% for 92% of the grids. Recent advances on the resolution and availability of remote sensing imagery, coupled with a decrease in its associated costs, have allowed the collection of timely information on soil and crop variability by examining spatial and temporal patterns of vegetation indices (Blackmer and White, 1996). Such information can be used to derive inputs for crop models in a GIS environment (Moran et al., 1997; Barnes et al., 1997). Vegetation indices, involving mathematical relationship and near infrared re¯ectance, have been extensively used with the goal of estimating vegetation amount (Wiegand et al., 1991; Jackson and Huete, 1991; Price, 1992). Among vegetation indices, the normalized dierence vegetative index (NDVI) is the one most commonly used to quantify canopy vigor and density (Price, 1992; Carlson and Ripley, 1997). NDVI is de®ned as: NDVI
NIR RED NIR RED
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
99
where NIR and RED represent the surface re¯ectance averaged over ranges of wavelengths in the near infrared (l 0.8 mm), and the visible (l 0.6 mm) regions of the spectrum, respectively. NDVI increases almost linearly with increasing leaf area index (LAI, leaf area per unit land area) until LAI exceeds values of 3±4, above which NDVI rapidly approaches an asymptotic limit (Liu and Huete, 1995; Jasinski 1996; Carlson and Ripley, 1997). NDVI analysis performed on images taken at critical times during a growing season can help characterize spatial variability in crop performance. Clearly, the goal of crop simulation in precision agriculture is to explain the spatial variability of crop performance mapped with grain yield monitoring systems and to help guide in management decisions related to the site-speci®c management of crop inputs. It is also clear that crop simulations cannot be performed everywhere given that the cost and the availability of detailed inputs would be prohibitive. A more balanced approach to the application of crop simulation models in precision agriculture would be to delineate zones within the ®eld of similar crop performance. One approach may be to obtain vegetation indexes derived from remote sensed imagery during critical times during the growing season, classify the images for target sampling, delineate spatial patterns and use the results of the target sampling as inputs for the models. Model validation can be then be performed at selected sites within these delineated management zones. Such approach would facilitate the challenge of using crop models in precision agriculture by obtaining spatial inputs to simulate variations of crop yields across the ®eld, as well as to decide where to use ®eld averages for some factors along with spatially variable inputs for others. The objective of this study was to examine a new procedure for spatial validation of crop models for use in precision agriculture that uses the CROPGRO-Soybean model to simulate soybean performance using variations in input aggregation. The procedure also uses the crop model to validate management zones across the ®eld delineated using a NDVI classi®cation procedure. 2. Materials and methods 2.1. Site description and ®eld measurements The study area consisted of a 7 ha portion of ®eld located 10 km south of Durand, MI. The ®eld has been cropped to a corn±soybean rotation for more than 10 years. Soils are variable containing ®ve soil map units and considerable spatial variability within soil fertility (Pierce et al., 1995; Pierce and Warncke, 2000) with the major soil type in the experimental area mapped as Capac loam (Udollic Ochraqualf ®ne, loamy, mixed, mesic). Soybean was grown in 1997 following Corn. The ®eld was planted on May 5 by direct drilling soybean (Variety Asgrow 1901, a Roundup Ready variety) in 37-cm rows at a seeding rate of 494,000 seeds ha 1. A regular grid consisting of 52 grid locations spaced 30.5 m apart was imposed on the 7 ha experimental area after planting. Position and elevation of each grid points were determined with a high-resolution dierential GPS. Neutron probe access
100
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
tubes were installed at each of the 52 grid locations. A neutron moisture gauge was used to measure soil water content at 15-cm increments to the depth of the C horizon or a maximum of 150 cm depth. Measurements were taken on a weekly basis throughout the season. During the installation of the neutron probe access tubes, soil samples were taken at the intersection of each of the 52 grid points in 25-cm increments and stored for analysis. Soil samples were air-dried and passed through a 2-mm sieve. Particle size was determined for each segment of each soil pro®le using the hydrometer method (Gee and Bauder, 1986). Soil organic matter was determined on the surface 25 cm of each soil pro®le by dry combustion using a CHN analyzer (Carlo Erba Instruments, Italy). The upper and lower limit of soil water availability was determined using soil water measurements taken in the ®eld, and from empirical equations based on soil texture (Ritchie and Crum, 1989; Ritchie et al., 1999). Soil depth for each grid point was determined using the deepest depth observed during the installation of neutron probe access tubes. Potential extractable soil water (PESW) was determined by subtracting the lower limit of plant water availability from upper limit for each soil layer and integrating it for the entire pro®le. A 5 m2 area was delineated at each grid location for selected plant measurements. Plant population and the distance between plants were measured at emergence (20 May) and at the third leaf stage of development (15 June). A non-destructive optical device with a ®sh-eye len sensor (LAI-2000, LI-COR) was used to quantify LAI of the 52 grid points on July 15 and August 10. Soybean yield was obtained by harvesting four rows along a 20-m length centered on each grid point using a plot combine. Grain moisture was obtained after harvest on a subsample from each harvested area. A datalogger (Licor 1000) was installed to collect weather data on solar radiation, minimum, maximum and mean temperature and precipitation, which are required as model input. Precipitation was measured with an electronic tipping bucket rain gauge every 5 min to record rainfall intensities as well as daily total amounts. Standard statistical analyses were performed for the variables measured in the ®eld. The spatial structure for each parameter was assessed thorough a semivariance analysis. Measurements taken on each grid point were interpolated using punctual kriging technique available in GS+ Version 3.1a (Gamma Design Software, 1999). Correlation matrices were developed to describe the relationships among variables for each single class and for the 52 grid points. 2.2. Remote sensing data The airborne false color composite images in the blue, red, green and NIR portion of the spectrum were taken on June 1, June 28, July 18, July 29, August 13, August 29, September 15 at 1-m pixel resolution. The images provided spatial information about the condition of the crop throughout the season. Each image was used to generate NDVI maps of the ®eld and to identify spatial patterns across the ®eld. The false color composite image taken on July 18 was selected for quantifying areas with similar re¯ectance by grouping areas into classes of similar NDVI values using supervised classi®cation technique available in Idrisi v32 software (Clark Labs,
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
101
1999). Pixels of similar re¯ectance were queried across the ®eld after trying various ranges of values able to reproduce the spatial patterns visible in the original falsecolor composite image. 2.3. Crop growth model CROPGRO-Soybean v.3.5 is a process oriented model that simulates plant responses to environmental conditions (soil and weather), genetics and management strategies. This model is part of the Decision Support System for Agrotechnolgy Transfer (DSSAT 3.5, Hoogenboom et al., 1994) that provides several tools for model application. The soil water limits used to run the simulation experiments varied spatially and according to the observed data of soil texture and soil water content at the 52 grid points. The model performance was evaluated using the Root Mean Square Error (RMSE): "
1X RMSE yi n i1
#1=2 y^ i
2
where yi are the measurements, yÃi the predictions, and n is the number of comparisons. 2.4. Simulation experiments The soybean model was used to simulate yields in the ®eld using inputs varying from ®eld averages to spatially variable inputs. Such approach was possible due to the intensive measurements of model inputs. Yields predictions were made using ®ve dierent scenarios. These scenarios varied from one that assumed uniform soil and management conditions across the ®eld to one that used ®eld-measured spatially variable inputs for the soil water balance parameters (LL, DUL, SAT and soil depth) and plant populations, to one that simulated three areas identi®ed by the NDVI analysis. The ®ve cases are described in Table 1.
Table 1 Scenarios for simulation experiments and number of model runs for each scenario Scenario No.
Model input
Model runs
RMSEa (kg ha 1)
1 2 3 4 5
Average of grids (Average soil type and target plant population) Average soil type and grid plant population Grid soil type and target plant population Grid soil type and grid plant population Average soil and plant population for 3 NDVI Classes
1 52 52 52 3
465 296 245 198 101
a
RMSE, root mean square error.
102
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
3. Results and discussion 3.1. Field measurements Soybean yield was spatially variable across the ®eld (Fig. 1a), ranging from 1900 to 3600 kg ha 1 with a mean value of 2953 kg ha 1 and a standard deviation of 433 kg ha 1. The spatial distribution of yield was consistent with other ®eld measurements (LAI and plant population) and by the remote sensing image that showed high re¯ectance in the high yielding areas. Field measurements of LAI on August 8 (Fig. 1b) re¯ected the dierent soil water regimes present across the ®eld. The highest LAI of 4.6 was observed in areas of high plant population, deeper soil, and high PESW. The areas of the ®eld with rocky soil and highest elevation had the lowest LAI of 1.7. The mean and the standard deviation for LAI were 3.6 and 0.6, respectively. The areas showing high LAI corresponded with the remote sensing image areas with high re¯ectance. Record cool weather in May delayed soybean emergence and resulted in variable population densities across the ®eld (Fig. 1c). Plant populations varied from 22 to 60 plants m 2 with a mean value for the 52 grid-points of 47 plants m 2 and a standard deviation of 8 plants m 2. Plant stand was highly in¯uenced by the environmental condition (soil and weather) at planting time. Soil textural analysis from the 52 grid-points showed high spatial variability for sand and clay particles. The clay content varied across the ®eld from 8% in high relative elevation areas to 25% in low relative elevation areas. Sand percentage varied from a minimum of 40% to a maximum of 82% and logically had an opposite spatial distribution from the clay content. Based on textural analysis results and their spatial distribution, three main soil types were detected across the ®eld. A deep-dark sandy-loam soil located in lower elevation areas of the ®eld, a sandy loam characterizing the ¯at areas, and a sandy-rocky soil present in the higher elevation areas. Soil depth measurements also showed the presence of high spatial structure across the ®eld (Fig. 2a). Peaks had lower soil depth due to erosion that depth to the C horizon. Soil depths ranged from 95 to 150 cm with a mean value for the 52 gridpoints of 130 cm and a standard deviation of 14 cm. Potential extractable soil water (PESW) was function of soil depth and texture, thus the spatial distribution of these variables were similar (Fig. 2b). A maximum PESW of 140 mm was observed in the low elevation areas, while the lowest value of 70 mm was found on peaks. The mean PESW value for the ®eld was 111 mm with a standard deviation of 19 mm. The ®eld is characterized by a rolling terrain that caused high spatial variability of soil properties. Landscape position and relative elevation highly in¯uenced soil physical properties thorough erosion processes that occurred over the years. The reclassi®ed NDVI map from 18 July image clearly showed spatial variability in soybean performance (Fig. 3). Classi®cation of the NDVI image indicated three classes of importance in this ®eld. Note that the areas of dierent classes are
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
103
Fig. 1. (a) Kriged map of measured soybean yield. (b) Kriged map of maximum LAI measured on August 8. (c) Kriged map of plant population measured on June 5.
not contiguous. Standard statistical analyses for the three NDVI classes and for the ®eld average are shown in Table 2. The ®eld values for each variable were obtained averaging measurements taken on the 52 grid points. The standard deviation for the
104
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Fig. 2. (a) Kriged map of measured soil depth. (b) Kriged map of plant extractable soil water.
®eld values was always higher compare with the NDVI classes except for elevation and soil depth. The correlation analysis between measured variables revealed that yield was highly correlated to LAI, PESW and NDVI as shown by the correlation coecients of 0.86, 0.87 and 0.94, respectively (Table 3). Positive correlation was also found between yield and plant population, and NDVI with LAI and plant population. Fig. 4 indicated the usefulness of NDVI as yield predictor and the importance of the appropriate timing to take a remote sensing image due to the NDVI properties (Carlson and Ripley, 1997). The spatial dependence was determined for each soil and crop variable measured in the ®eld. Geostatistical analysis revealed spatial structure for all the variables giving ranges of distance that varied from 60 m for the plant population to 150 m
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
105
Fig. 3. Reclassi®ed NDVI image showing the three NDVI classes identi®ed.
for the yield (Table 4). The semivariance of the measured variables was ®tted well using a spherical model. 3.2. Crop simulations 3.2.1. Scenarios 1±4 Error in yield prediction decreased as the level of input detail increased for the simulation scenarios tested (Table 1). The ®eld average of 2950 kg ha 1 was underestimated under Scenario 1, which predicted a soybean yield of 2530 kg ha 1. The RMSE for Scenario 1 was 465 kg ha 1. Under Scenario 2, adding site-speci®c plant population data as model input improved model performance by decreasing the RMSE to 296 kg ha 1, a reduction of 36% over Scenario 1. Using site-speci®c soils data at constant plant populations in Scenario 3 improved yield prediction 18% more than Scenario 2 as evidenced by an RMSE of 245 kg ha 1, an improvement of 47% over Scenario 1. Using both site-speci®c soils and plant population input further reduced RMSE improving the prediction of soybean yield over Scenario 1 by 58%. 3.2.2. NDVI classes (Scenario 5) Fig. 5a and b show the kriged map of measured and simulated yield for the three NDVI classes. The model was able to simulate well the yield for the three NDVI classes as evidenced by RMSE values for the three classes (Table 5) that were lower than the other scenarios. The model performance indicated that the NDVI reclassi®cation procedure was appropriate and with multi-year simulation should allow for the characterization of management zones for this ®eld. The progressive use of site speci®c model inputs combined with the NDVIreclassi®cation has a major advantage since the power and application of simulation models in precision agriculture has been limited by scale. The site-speci®c inputs approach is scale-independent because the scale is controlled by the observed variation in the ®eld and that is the scale at which the model is applied.
106
Class ID
Cell NDVI LAI Plant population (pl m 2) counts class range Mean S.D. Mean S.D.
1 3159 0±0.25 2 42 528 0.25±0.5 3 23 916 >0.5 52 Grid 69 603 ± point a
2.16 3.24 3.95 3.6
0.41 0.44 0.38 0.6
36 42 50 47
13 7 7 8
Yield (kg ha 1)
Elevation (m)
Soil depth (cm)
PESW (mm)
NDVI
Mean
S.D.
Mean
S.D.
Mean
S.D.
Mean
S.D.
Mean S.D.
1635 2729 3155 2953
339 217 183 433
263.6 262.7 262.4 262.5
0.8 0.4 0.7 0.7
110 128 134 130
22 7 13 14
73 94 120 111
6 9 12 19
0.10 0.43 0.59 0.51
NDVI, normalized dierence vegatative index; LAI, leaf area index; PESW, potential extractable soil water.
0.03 0.05 0.05 0.14
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Table 2 Standard statistical analysis for the variables measured in each NDVI class and average for the ®eld (52 grid points)a
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
107
Table 3 Correlation matrix among variables measured in the ®eld at the 52 grid-pointsa Variables
LAI
LAI Plant population Yield Elevation Soil depth % Clay (0-75 cm) % Sand (0-75cm) PESW NDVI
1 0.76 0.86 0.42 0.32 0.05 0.18 0.82 0.87
Plant Yield population 1 0.68 0.44 0.26 0.07 0.18 0.64 0.65
1 0.49 0.31 0.03 0.13 0.87 0.94
Elevation Soil % Clay % Sand PESW NDVI depth (0±75 cm) (0±75 cm)
1 0.18 0.15 0.14 0.46 0.48
1 0.17 0.23 0.32 0.29
1 0.81 0.19 0.09
1 0.34 0.18
1 0.88
1
a LAI, leaf area index; PESW, potential extractable soil water; NDVI, normalized dierence vegatative index.
Fig. 4. Correlation between NDVI (image taken on July 18) and measured yield.
4. Conclusion Remote sensing imagery with NDVI analysis allowed for the identi®cation of spatial patterns of crop growth variability. The variability in soybean populations within the ®eld provided validation of the plant population and soil type eects on soybean yields predicted by the CROPGRO-Soybean model. The model was able to
108
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Table 4 Semivariograms parameters for the variables measured in the ®elda Variables
Model
Nugget
Sill
Range
Yield LAI Plant Population Soil Depth Clay PESW
Spherical Spherical Spherical Spherical Spherical Spherical
0.295 0.279 0.110 0.202 0.210 0.281
1.294 1.156 1.094 1.149 1.095 1.107
150 120 60 100 110 100
a
LAI, leaf area index; PESW, potential extractable soil water.
Fig. 5. (a) Kriged map measured yield for the three NDVI classes. (b) Kriged map of simulated yield for the three NDVI classes using average measured inputs within the each class.
Simulation experiments
Number of grid point
Total area (ha)
Measured yield (kg ha 1)
Simulated yield (kg ha 1)
RMSE (kg ha 1)
Average Inputs within NDVI-Class 1 Average Inputs within NDVI-Class 2 Average Inputs within NDVI-Class 3 Weighted Average Inputs for all NDVI Classes (Scenario 5)
7 15 30 52
0.32 4.25 2.39 6.96
1199 2785 3087 2849
1216 2802 3203 2950
17 17 116 101
a
RMSE, root mean square error; NDVI, normalized dierence vegatative index.
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Table 5 Summary table for measured and simulated yield for each NDVI Classa
109
110
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
closely predict the yield across the ®eld when the correct inputs were used, showing great potential for use in yield map prediction and interpretation in the contest of precision agriculture. It is clear that zone-speci®c management to optimize production can be developed through a combination of remote sensing and crop simulation models. This is a more aordable alternative to the use of traditional soil sampling and micro-scale sensor. It also answers questions related to scale issue by applying the model at the scale of observed variability through remote sensing and NDVI image reclassi®cation.
Acknowledgements The authors wish to express their gratitude to the Michigan Soybean Promotion Committee and to the United Soybean Board for funding this project. A special thanks goes to John Anibal, Brian Long, Cal Bricker for their valuable help in the ®eld data collection.
References Barnes, E.M., Pinter Jr., P.J., Kimball, B.A., Wall, G.W., LaMorte, R.L., Hunsaker, D.J., Adamsen, F., Leavitt, S., Thompson, T., Mathius, J., 1997. Modi®cation of CERES-Wheat to accept leaf area index as an input variable. Written for presentation at the 1997 ASAE Annual International Meeting, sponsored by ASAE, Minneapolis, MN, August 10±14, 1997. (Paper No. 9730160). Batchelor, W.D., Paz, J.O., 1998. Process-oriented crop growth models as a tool to evaluate spatial yield variability. In: Proceedings of the First International Conference on Geospatial Information in Agriculture and Forestry, 1±3 June 1998, Vol. 1. ERIM International, Lake Buena Vista, Florida, USA, pp. I198±205. Blackmer, A.M., White, S.E., 1996. Remote sensing to identify spatial patterns in optimal rates of nitrogen fertilization. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Proceedings of the Third International Conference on Precision Agriculture, Minneapolis, MN, 23±26 June 1996. ASA Miscellaneous Publication ASA, CSSA, and SSSA, Madison, WI, pp. 33±41. Boote, K.J., Jones, J.W., Hoogenboom, G., Pickering, N.B., 1998. CROPGRO model for grain legumes. In: Tsuji, G.Y., Hoogenboom, G., Thornton, P.K. (Eds.), Understanding Options for Agricultural Production. Kluwer Academic Publishers, Dordrecht, Netherlands, pp. 99±128. Cambardella, C.A., Colvin, T.S., Karlen, D.L., Logsdon, S.D., Berry, E.C., Radke, T.C., Kaspar, T.B., Parkin, X., Jaynes, D.B., 1996. Soil property contributions to yield variation pattern. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Proceedings of the Third International Conference on Precision Agriculture. ASA, CSSA, SSSA, Inc, Madison, WI, pp. 417±424. Carlson, T.N., Ripley, D.A., 1997. On the relation between NDVI, fractional vegetation cover, and leaf area index. Remote Sens. Environ. 62, 241±252. Cora, J.E., Pierce, F.J., Basso, B., Ritchie, J.T., 1999. Simulation of within-®eld variability of corn yield with Ceres-Maize model. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Precision Agriculture. ASA-CSSA-SSSA, Madison, Wisconsin, USA, pp. 1309±1319. Gee, G.W., Bauder, J.W., 1986. Particle-size analysis. In: Methods of Soil and Plant Analysis I. American Society of Agronomy, Madison, Wisconsin, USA, pp. 383±411. Gamma Design Software, 1999. Gs+: geostatistics for the environmental scienc. Version 3.5. Gamma Design Software, Plainwell, MI, p. 44.
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
111
Hoogenboom, G.J., Jones, J.W., Wilkens, P.W., Batchelor, W.D., Bowen, W.T., Hunt, L.A. et al., 1994. Crop models. In: Tsuji, G.Y., Uehara, G., Balas, S. (Eds.), DSSAT v3., Vol. 2-2. University of Hawaii, Honolulu, Hawaii. Jackson, R.D., Huete, A.R., 1991. Interpreting vegetation indeces. Preventive Veterinary Medicine 11, 185±200. Jasinski, M.F., 1996. Estimation of subpixel vegetation density of natural regions using satellite multispectral imagery. IEEE Trans. Geosci. Remote Sens. 34, 804±813. Jones, J.W., Ritchie, J.T., 1991. Crop growth models. In: Homan, G.J., Howell, T.A., Soloman, K.H. (Eds.), Management of Farm Irrigation Systems. ASAE, St. Joseph, MI. Liu, Q., Huete, A., 1995. A feedback based modi®cation of the NDVI to minimize canopy background and atmosphere noise. IEEE Trans. Geosci. Remote Sens. 33, 457±465. Moran, M.S., Inoue, Y., Barnes, E.M., 1997. Opportunities and limitations for image-based remote sensing in precision crop management. Remote Sensing of the Environment 61, 319±346. Paz, J.O., Batchelor, W.D., Colvin, T.S., Logsdon, S.D., Kaspar, T.C., Karlen, D.L., 1997. Analysis of water stress eects causing spatial yield variability in soybeans. Trans. ASAE 41, 1527±1534. Paz, J.O., Batchelor, W.D., Babcock, B.A., Colvin, T.S., Logsdon, S.D., Kaspar, T.C., Karlen, D.L., 1999. Model-based technique to determine variable rate nitrogen for corn. Agric. Syst. 61, 69±75. Pierce, F.J., Nowak, P., 1999. Aspects of precision agriculture. Advances in Agronomy 67, 1±86. Pierce, F.J., Warncke, D.D., 2000. Soil and crop response to variable-rate limiting for two Michigan ®elds. Soil Sci. Soc. Am J. 64, 774±780. Pierce, F.J., Warncke, D.D., Everett, M.Wm, 1995. Yield and nutrient availability in glacial soils of Michigan. In: Robert, P.C. et al. (Eds.), Proc. of the Int. Conf. on Site Speci®c Management for Agricultural Systems, 2nd. Bloomington, WI. Pierce, F.J., Anderson, N.W., Colvin, T.S., Schueller, J.K., Humberg, D.S., Mc.Laughlin, N.B., 1997. Yield mapping. In: Pierce, F.J., Sadler, E.J. (Eds.), The State of Site Speci®c Management for Agriculture. ASA Misc. Publ., ASA, CSSA, and SSSA, Madison, WI, pp. 211±243. Price, J.C., 1992. Estimating vegetation amount from visible and near infrared re¯ectances. Remote Sens. Environ. 41, 29±34. Ritchie, J.T., 1987. Using crop models as a decision support system to reduce nitrate leaching. In: D'Itri, F.M., Wolfson, L.G. (Eds.), Rural Groundwater Contamination. Lewis Publishers, Chelsea, Michigan, pp. 179±192. Ritchie, J.T., 1991. Speci®cations of the ideal model for predicting crop yield. In: Muchow, R.C., Bellamy, J.A. (Eds.), Climatic Risk in Crop Production: Model and Management for the Semi-arid Tropic and Subtropic. Proc. Intl. Symp. S. Lucia, Brisbane, Queensland, Australia 2±6 July, 1990. CAB International, Wallingford, U.K., pp. 97±122. Ritchie, J.T., Crum, J., 1989. Converting soil survey characterization data into IBSNAT crop model input. In: Bouma, J., Berget, A.K. (Eds.), Land Qualities in Space and Time. Wageningen, the Netherlands, pp. 155±167. Ritchie, J.T., Godwin, D.C., Otter-Nacke, S., 1985. CERES-Wheat: a simulation model of wheat growth and development. Texas A&M University Press, College Station, TX. Ritchie, J.T., Gerakis, A., Suleiman, A., 1999. Simple model to estimate ®eld-measured soil water limits. Tran. ASAE 42, 1609±1614. Sadler, E.J., Russell, G., 1997. Modeling crop yield for site speci®c management. In: Pierce, F.J., Sadler, E.J. (Eds.), The State of Site Speci®c Management for Agriculture. ASA Misc. Publ., ASA, CSSA, and SSSA, Madison, WI, pp. 69±80. Sadler, E.J., Bauer, P.J., Busscher, W.J., 2000. Site-speci®c analysis of a droughted corn crop: I. growth and grain yield. Agron. J. 92, 395±402. Sudduth, K.A., Drummond, S.T., Birrell, S.J., Kitchen, N.R., 1996. Analysis of spatial factors in¯uencing crop yield. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Proceedings of the Third International Conference on Precision Agriculture. ASA, CSSA, SSSA, Madison, WI, pp. 129±140. Verhagen, J., Bouma, J., 1997. Modeling soil variability. In: Pierce, F.J., Sadler, E.J. (Eds.), The State of
112
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Site Speci®c Management for Agriculture. ASA Misc. Publ., ASA, CSSA, and SSSA, Madison, WI, pp. 55±68. Verhagen, A., Booltink, H.W.G., Bouma, J., 1995. Site-speci®c management: balancing production and environmental requirements at farm scale. Agric. Syst. 49, 369±384. Wiegand, C.L., Richardson, A.J., Escobar, D.E., Gerbermann, A.H., 1991. Vegetation indeces in crop assessments. Remote Sens. Environ. 35, 105±119.
www.elsevier.com/locate/agsy
Spatial validation of crop models for precision agriculture B. Basso a,*, J.T. Ritchie b, F.J. Pierce c, R.P. Braga d, J.W. Jones d a Dipartimento di Produzione Vegetale, Universita' degli Studi della Basilicata, 85100 Potenza, Italy Department of Crop and Soil Sciences, Michigan State University, East Lansing, MI 48824-1325, USA c Center for Precision Agricultural Systems, Washington State University Prosser, WA 99350, USA d Department of Agricultural Engineering, University of Florida, Gainsville, FL 32611, USA
b
Received 31 July 2000; received in revised form 11 October 2000; accepted 24 October 2000
Abstract Spatial measurements of yield using technological advances like on-the-go yield monitoring systems have clearly shown large within-®eld variability in crop yields suggesting that ®eld yields could be increased or cost decreased by varying management over space. This study evaluated the utility of the CROPGRO-Soybean simulation model and remote sensing in the interpretation of a soybean yield map. CROPGRO was executed on areas within the ®eld de®ned as reasonably uniform by a Normalized Dierence Vegetative Index (NDVI) analysis. The model was able to closely predict the crop yield variability measured within the ®eld when the measured soil type and plant population were used as model inputs. Remote sensing was useful in ®nding spatial patterns across the ®eld to target sampling and to provide spatial inputs for the model. Results of this study showed that a combination of crop model and remote sensing can identify management zones and causes for yield variability, which are prerequisites for zone-speci®c management prescriptions. # 2001 Elsevier Science Ltd. All rights reserved. Keywords: Crop yield spatial variability; Crop models; Remote sensing; NDVI; Management zones
* Corresponding author. Tel.: +39-971-202-271; fax: +39-971-202-269. E-mail address: [email protected] (B. Basso). 0308-521X/01/$ - see front matter # 2001 Elsevier Science Ltd. All rights reserved. PII: S0308-521X(00)00063-9
98
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
1. Introduction Agricultural production systems are inherently variable due to spatial variation in soil properties, topography, climate and other factors. To achieve the ultimate goal of sustainable cropping systems, variability must be considered both in space and time because the factors in¯uencing crop yield have dierent spatial and temporal behavior (Verghagen and Bouma, 1997; Pierce and Nowak, 1999). Advances in technologies such as Global Positioning Systems (GPS), Geographic Information Systems (GIS) and remote sensing have created the possibility to assess the spatial variability present in the ®eld and manage it with appropriate site-speci®c practices (Pierce et al., 1997; Verhagen et al., 1995). Site-speci®c management (SSM) strategies may be able to optimize production, but their potential bene®ts are highly dependent on the accuracy of the assessment of such variability (Pierce and Nowak, 1999). Traditional analytical techniques, such as regression of static measurements against yield, have failed to explain the reasons for yield variability because the dynamic, thus temporal, interactions of stresses on crop growth and development cannot be accounted for (Jones and Ritchie, 1991; Cambardella et al., 1996; Sudduth et al., 1996). Process oriented crop simulation models, such as the CERES and CROPGRO (Ritchie et al., 1985; Boote et al., 1998), integrate the eects of temporal and multiple stress interactions on crop growth processes under dierent environmental and management conditions. Even though crop models have shown high potential for optimizing production and minimizing environmental impact (Ritchie, 1987, 1991), their application for SSM has been limited thus far (Sadler et al., 2000). Crop models can be used for understanding yield variability in both space and time, leading to a more sustainable environment (Sadler and Russell, 1997; Cora et al., 1999). Batchelor et al. (1998) and Paz et al. (1997, 1999) used CERES-Maize and CROPGRO-Soybean simulation models to determine the eect of soil moisture variation throughout the season on yield spatial variability optimizing for soil water limit parameters. The differences between measured and predicted yield for 224 grid points over a 3-year period ranged from 10% for 70% of the grids and 20% for 96% of the grids in maize and for soybean from 10% for 84% of the grids and 20% for 92% of the grids. Recent advances on the resolution and availability of remote sensing imagery, coupled with a decrease in its associated costs, have allowed the collection of timely information on soil and crop variability by examining spatial and temporal patterns of vegetation indices (Blackmer and White, 1996). Such information can be used to derive inputs for crop models in a GIS environment (Moran et al., 1997; Barnes et al., 1997). Vegetation indices, involving mathematical relationship and near infrared re¯ectance, have been extensively used with the goal of estimating vegetation amount (Wiegand et al., 1991; Jackson and Huete, 1991; Price, 1992). Among vegetation indices, the normalized dierence vegetative index (NDVI) is the one most commonly used to quantify canopy vigor and density (Price, 1992; Carlson and Ripley, 1997). NDVI is de®ned as: NDVI
NIR RED NIR RED
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
99
where NIR and RED represent the surface re¯ectance averaged over ranges of wavelengths in the near infrared (l 0.8 mm), and the visible (l 0.6 mm) regions of the spectrum, respectively. NDVI increases almost linearly with increasing leaf area index (LAI, leaf area per unit land area) until LAI exceeds values of 3±4, above which NDVI rapidly approaches an asymptotic limit (Liu and Huete, 1995; Jasinski 1996; Carlson and Ripley, 1997). NDVI analysis performed on images taken at critical times during a growing season can help characterize spatial variability in crop performance. Clearly, the goal of crop simulation in precision agriculture is to explain the spatial variability of crop performance mapped with grain yield monitoring systems and to help guide in management decisions related to the site-speci®c management of crop inputs. It is also clear that crop simulations cannot be performed everywhere given that the cost and the availability of detailed inputs would be prohibitive. A more balanced approach to the application of crop simulation models in precision agriculture would be to delineate zones within the ®eld of similar crop performance. One approach may be to obtain vegetation indexes derived from remote sensed imagery during critical times during the growing season, classify the images for target sampling, delineate spatial patterns and use the results of the target sampling as inputs for the models. Model validation can be then be performed at selected sites within these delineated management zones. Such approach would facilitate the challenge of using crop models in precision agriculture by obtaining spatial inputs to simulate variations of crop yields across the ®eld, as well as to decide where to use ®eld averages for some factors along with spatially variable inputs for others. The objective of this study was to examine a new procedure for spatial validation of crop models for use in precision agriculture that uses the CROPGRO-Soybean model to simulate soybean performance using variations in input aggregation. The procedure also uses the crop model to validate management zones across the ®eld delineated using a NDVI classi®cation procedure. 2. Materials and methods 2.1. Site description and ®eld measurements The study area consisted of a 7 ha portion of ®eld located 10 km south of Durand, MI. The ®eld has been cropped to a corn±soybean rotation for more than 10 years. Soils are variable containing ®ve soil map units and considerable spatial variability within soil fertility (Pierce et al., 1995; Pierce and Warncke, 2000) with the major soil type in the experimental area mapped as Capac loam (Udollic Ochraqualf ®ne, loamy, mixed, mesic). Soybean was grown in 1997 following Corn. The ®eld was planted on May 5 by direct drilling soybean (Variety Asgrow 1901, a Roundup Ready variety) in 37-cm rows at a seeding rate of 494,000 seeds ha 1. A regular grid consisting of 52 grid locations spaced 30.5 m apart was imposed on the 7 ha experimental area after planting. Position and elevation of each grid points were determined with a high-resolution dierential GPS. Neutron probe access
100
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
tubes were installed at each of the 52 grid locations. A neutron moisture gauge was used to measure soil water content at 15-cm increments to the depth of the C horizon or a maximum of 150 cm depth. Measurements were taken on a weekly basis throughout the season. During the installation of the neutron probe access tubes, soil samples were taken at the intersection of each of the 52 grid points in 25-cm increments and stored for analysis. Soil samples were air-dried and passed through a 2-mm sieve. Particle size was determined for each segment of each soil pro®le using the hydrometer method (Gee and Bauder, 1986). Soil organic matter was determined on the surface 25 cm of each soil pro®le by dry combustion using a CHN analyzer (Carlo Erba Instruments, Italy). The upper and lower limit of soil water availability was determined using soil water measurements taken in the ®eld, and from empirical equations based on soil texture (Ritchie and Crum, 1989; Ritchie et al., 1999). Soil depth for each grid point was determined using the deepest depth observed during the installation of neutron probe access tubes. Potential extractable soil water (PESW) was determined by subtracting the lower limit of plant water availability from upper limit for each soil layer and integrating it for the entire pro®le. A 5 m2 area was delineated at each grid location for selected plant measurements. Plant population and the distance between plants were measured at emergence (20 May) and at the third leaf stage of development (15 June). A non-destructive optical device with a ®sh-eye len sensor (LAI-2000, LI-COR) was used to quantify LAI of the 52 grid points on July 15 and August 10. Soybean yield was obtained by harvesting four rows along a 20-m length centered on each grid point using a plot combine. Grain moisture was obtained after harvest on a subsample from each harvested area. A datalogger (Licor 1000) was installed to collect weather data on solar radiation, minimum, maximum and mean temperature and precipitation, which are required as model input. Precipitation was measured with an electronic tipping bucket rain gauge every 5 min to record rainfall intensities as well as daily total amounts. Standard statistical analyses were performed for the variables measured in the ®eld. The spatial structure for each parameter was assessed thorough a semivariance analysis. Measurements taken on each grid point were interpolated using punctual kriging technique available in GS+ Version 3.1a (Gamma Design Software, 1999). Correlation matrices were developed to describe the relationships among variables for each single class and for the 52 grid points. 2.2. Remote sensing data The airborne false color composite images in the blue, red, green and NIR portion of the spectrum were taken on June 1, June 28, July 18, July 29, August 13, August 29, September 15 at 1-m pixel resolution. The images provided spatial information about the condition of the crop throughout the season. Each image was used to generate NDVI maps of the ®eld and to identify spatial patterns across the ®eld. The false color composite image taken on July 18 was selected for quantifying areas with similar re¯ectance by grouping areas into classes of similar NDVI values using supervised classi®cation technique available in Idrisi v32 software (Clark Labs,
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
101
1999). Pixels of similar re¯ectance were queried across the ®eld after trying various ranges of values able to reproduce the spatial patterns visible in the original falsecolor composite image. 2.3. Crop growth model CROPGRO-Soybean v.3.5 is a process oriented model that simulates plant responses to environmental conditions (soil and weather), genetics and management strategies. This model is part of the Decision Support System for Agrotechnolgy Transfer (DSSAT 3.5, Hoogenboom et al., 1994) that provides several tools for model application. The soil water limits used to run the simulation experiments varied spatially and according to the observed data of soil texture and soil water content at the 52 grid points. The model performance was evaluated using the Root Mean Square Error (RMSE): "
1X RMSE yi n i1
#1=2 y^ i
2
where yi are the measurements, yÃi the predictions, and n is the number of comparisons. 2.4. Simulation experiments The soybean model was used to simulate yields in the ®eld using inputs varying from ®eld averages to spatially variable inputs. Such approach was possible due to the intensive measurements of model inputs. Yields predictions were made using ®ve dierent scenarios. These scenarios varied from one that assumed uniform soil and management conditions across the ®eld to one that used ®eld-measured spatially variable inputs for the soil water balance parameters (LL, DUL, SAT and soil depth) and plant populations, to one that simulated three areas identi®ed by the NDVI analysis. The ®ve cases are described in Table 1.
Table 1 Scenarios for simulation experiments and number of model runs for each scenario Scenario No.
Model input
Model runs
RMSEa (kg ha 1)
1 2 3 4 5
Average of grids (Average soil type and target plant population) Average soil type and grid plant population Grid soil type and target plant population Grid soil type and grid plant population Average soil and plant population for 3 NDVI Classes
1 52 52 52 3
465 296 245 198 101
a
RMSE, root mean square error.
102
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
3. Results and discussion 3.1. Field measurements Soybean yield was spatially variable across the ®eld (Fig. 1a), ranging from 1900 to 3600 kg ha 1 with a mean value of 2953 kg ha 1 and a standard deviation of 433 kg ha 1. The spatial distribution of yield was consistent with other ®eld measurements (LAI and plant population) and by the remote sensing image that showed high re¯ectance in the high yielding areas. Field measurements of LAI on August 8 (Fig. 1b) re¯ected the dierent soil water regimes present across the ®eld. The highest LAI of 4.6 was observed in areas of high plant population, deeper soil, and high PESW. The areas of the ®eld with rocky soil and highest elevation had the lowest LAI of 1.7. The mean and the standard deviation for LAI were 3.6 and 0.6, respectively. The areas showing high LAI corresponded with the remote sensing image areas with high re¯ectance. Record cool weather in May delayed soybean emergence and resulted in variable population densities across the ®eld (Fig. 1c). Plant populations varied from 22 to 60 plants m 2 with a mean value for the 52 grid-points of 47 plants m 2 and a standard deviation of 8 plants m 2. Plant stand was highly in¯uenced by the environmental condition (soil and weather) at planting time. Soil textural analysis from the 52 grid-points showed high spatial variability for sand and clay particles. The clay content varied across the ®eld from 8% in high relative elevation areas to 25% in low relative elevation areas. Sand percentage varied from a minimum of 40% to a maximum of 82% and logically had an opposite spatial distribution from the clay content. Based on textural analysis results and their spatial distribution, three main soil types were detected across the ®eld. A deep-dark sandy-loam soil located in lower elevation areas of the ®eld, a sandy loam characterizing the ¯at areas, and a sandy-rocky soil present in the higher elevation areas. Soil depth measurements also showed the presence of high spatial structure across the ®eld (Fig. 2a). Peaks had lower soil depth due to erosion that depth to the C horizon. Soil depths ranged from 95 to 150 cm with a mean value for the 52 gridpoints of 130 cm and a standard deviation of 14 cm. Potential extractable soil water (PESW) was function of soil depth and texture, thus the spatial distribution of these variables were similar (Fig. 2b). A maximum PESW of 140 mm was observed in the low elevation areas, while the lowest value of 70 mm was found on peaks. The mean PESW value for the ®eld was 111 mm with a standard deviation of 19 mm. The ®eld is characterized by a rolling terrain that caused high spatial variability of soil properties. Landscape position and relative elevation highly in¯uenced soil physical properties thorough erosion processes that occurred over the years. The reclassi®ed NDVI map from 18 July image clearly showed spatial variability in soybean performance (Fig. 3). Classi®cation of the NDVI image indicated three classes of importance in this ®eld. Note that the areas of dierent classes are
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
103
Fig. 1. (a) Kriged map of measured soybean yield. (b) Kriged map of maximum LAI measured on August 8. (c) Kriged map of plant population measured on June 5.
not contiguous. Standard statistical analyses for the three NDVI classes and for the ®eld average are shown in Table 2. The ®eld values for each variable were obtained averaging measurements taken on the 52 grid points. The standard deviation for the
104
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Fig. 2. (a) Kriged map of measured soil depth. (b) Kriged map of plant extractable soil water.
®eld values was always higher compare with the NDVI classes except for elevation and soil depth. The correlation analysis between measured variables revealed that yield was highly correlated to LAI, PESW and NDVI as shown by the correlation coecients of 0.86, 0.87 and 0.94, respectively (Table 3). Positive correlation was also found between yield and plant population, and NDVI with LAI and plant population. Fig. 4 indicated the usefulness of NDVI as yield predictor and the importance of the appropriate timing to take a remote sensing image due to the NDVI properties (Carlson and Ripley, 1997). The spatial dependence was determined for each soil and crop variable measured in the ®eld. Geostatistical analysis revealed spatial structure for all the variables giving ranges of distance that varied from 60 m for the plant population to 150 m
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
105
Fig. 3. Reclassi®ed NDVI image showing the three NDVI classes identi®ed.
for the yield (Table 4). The semivariance of the measured variables was ®tted well using a spherical model. 3.2. Crop simulations 3.2.1. Scenarios 1±4 Error in yield prediction decreased as the level of input detail increased for the simulation scenarios tested (Table 1). The ®eld average of 2950 kg ha 1 was underestimated under Scenario 1, which predicted a soybean yield of 2530 kg ha 1. The RMSE for Scenario 1 was 465 kg ha 1. Under Scenario 2, adding site-speci®c plant population data as model input improved model performance by decreasing the RMSE to 296 kg ha 1, a reduction of 36% over Scenario 1. Using site-speci®c soils data at constant plant populations in Scenario 3 improved yield prediction 18% more than Scenario 2 as evidenced by an RMSE of 245 kg ha 1, an improvement of 47% over Scenario 1. Using both site-speci®c soils and plant population input further reduced RMSE improving the prediction of soybean yield over Scenario 1 by 58%. 3.2.2. NDVI classes (Scenario 5) Fig. 5a and b show the kriged map of measured and simulated yield for the three NDVI classes. The model was able to simulate well the yield for the three NDVI classes as evidenced by RMSE values for the three classes (Table 5) that were lower than the other scenarios. The model performance indicated that the NDVI reclassi®cation procedure was appropriate and with multi-year simulation should allow for the characterization of management zones for this ®eld. The progressive use of site speci®c model inputs combined with the NDVIreclassi®cation has a major advantage since the power and application of simulation models in precision agriculture has been limited by scale. The site-speci®c inputs approach is scale-independent because the scale is controlled by the observed variation in the ®eld and that is the scale at which the model is applied.
106
Class ID
Cell NDVI LAI Plant population (pl m 2) counts class range Mean S.D. Mean S.D.
1 3159 0±0.25 2 42 528 0.25±0.5 3 23 916 >0.5 52 Grid 69 603 ± point a
2.16 3.24 3.95 3.6
0.41 0.44 0.38 0.6
36 42 50 47
13 7 7 8
Yield (kg ha 1)
Elevation (m)
Soil depth (cm)
PESW (mm)
NDVI
Mean
S.D.
Mean
S.D.
Mean
S.D.
Mean
S.D.
Mean S.D.
1635 2729 3155 2953
339 217 183 433
263.6 262.7 262.4 262.5
0.8 0.4 0.7 0.7
110 128 134 130
22 7 13 14
73 94 120 111
6 9 12 19
0.10 0.43 0.59 0.51
NDVI, normalized dierence vegatative index; LAI, leaf area index; PESW, potential extractable soil water.
0.03 0.05 0.05 0.14
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Table 2 Standard statistical analysis for the variables measured in each NDVI class and average for the ®eld (52 grid points)a
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
107
Table 3 Correlation matrix among variables measured in the ®eld at the 52 grid-pointsa Variables
LAI
LAI Plant population Yield Elevation Soil depth % Clay (0-75 cm) % Sand (0-75cm) PESW NDVI
1 0.76 0.86 0.42 0.32 0.05 0.18 0.82 0.87
Plant Yield population 1 0.68 0.44 0.26 0.07 0.18 0.64 0.65
1 0.49 0.31 0.03 0.13 0.87 0.94
Elevation Soil % Clay % Sand PESW NDVI depth (0±75 cm) (0±75 cm)
1 0.18 0.15 0.14 0.46 0.48
1 0.17 0.23 0.32 0.29
1 0.81 0.19 0.09
1 0.34 0.18
1 0.88
1
a LAI, leaf area index; PESW, potential extractable soil water; NDVI, normalized dierence vegatative index.
Fig. 4. Correlation between NDVI (image taken on July 18) and measured yield.
4. Conclusion Remote sensing imagery with NDVI analysis allowed for the identi®cation of spatial patterns of crop growth variability. The variability in soybean populations within the ®eld provided validation of the plant population and soil type eects on soybean yields predicted by the CROPGRO-Soybean model. The model was able to
108
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Table 4 Semivariograms parameters for the variables measured in the ®elda Variables
Model
Nugget
Sill
Range
Yield LAI Plant Population Soil Depth Clay PESW
Spherical Spherical Spherical Spherical Spherical Spherical
0.295 0.279 0.110 0.202 0.210 0.281
1.294 1.156 1.094 1.149 1.095 1.107
150 120 60 100 110 100
a
LAI, leaf area index; PESW, potential extractable soil water.
Fig. 5. (a) Kriged map measured yield for the three NDVI classes. (b) Kriged map of simulated yield for the three NDVI classes using average measured inputs within the each class.
Simulation experiments
Number of grid point
Total area (ha)
Measured yield (kg ha 1)
Simulated yield (kg ha 1)
RMSE (kg ha 1)
Average Inputs within NDVI-Class 1 Average Inputs within NDVI-Class 2 Average Inputs within NDVI-Class 3 Weighted Average Inputs for all NDVI Classes (Scenario 5)
7 15 30 52
0.32 4.25 2.39 6.96
1199 2785 3087 2849
1216 2802 3203 2950
17 17 116 101
a
RMSE, root mean square error; NDVI, normalized dierence vegatative index.
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Table 5 Summary table for measured and simulated yield for each NDVI Classa
109
110
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
closely predict the yield across the ®eld when the correct inputs were used, showing great potential for use in yield map prediction and interpretation in the contest of precision agriculture. It is clear that zone-speci®c management to optimize production can be developed through a combination of remote sensing and crop simulation models. This is a more aordable alternative to the use of traditional soil sampling and micro-scale sensor. It also answers questions related to scale issue by applying the model at the scale of observed variability through remote sensing and NDVI image reclassi®cation.
Acknowledgements The authors wish to express their gratitude to the Michigan Soybean Promotion Committee and to the United Soybean Board for funding this project. A special thanks goes to John Anibal, Brian Long, Cal Bricker for their valuable help in the ®eld data collection.
References Barnes, E.M., Pinter Jr., P.J., Kimball, B.A., Wall, G.W., LaMorte, R.L., Hunsaker, D.J., Adamsen, F., Leavitt, S., Thompson, T., Mathius, J., 1997. Modi®cation of CERES-Wheat to accept leaf area index as an input variable. Written for presentation at the 1997 ASAE Annual International Meeting, sponsored by ASAE, Minneapolis, MN, August 10±14, 1997. (Paper No. 9730160). Batchelor, W.D., Paz, J.O., 1998. Process-oriented crop growth models as a tool to evaluate spatial yield variability. In: Proceedings of the First International Conference on Geospatial Information in Agriculture and Forestry, 1±3 June 1998, Vol. 1. ERIM International, Lake Buena Vista, Florida, USA, pp. I198±205. Blackmer, A.M., White, S.E., 1996. Remote sensing to identify spatial patterns in optimal rates of nitrogen fertilization. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Proceedings of the Third International Conference on Precision Agriculture, Minneapolis, MN, 23±26 June 1996. ASA Miscellaneous Publication ASA, CSSA, and SSSA, Madison, WI, pp. 33±41. Boote, K.J., Jones, J.W., Hoogenboom, G., Pickering, N.B., 1998. CROPGRO model for grain legumes. In: Tsuji, G.Y., Hoogenboom, G., Thornton, P.K. (Eds.), Understanding Options for Agricultural Production. Kluwer Academic Publishers, Dordrecht, Netherlands, pp. 99±128. Cambardella, C.A., Colvin, T.S., Karlen, D.L., Logsdon, S.D., Berry, E.C., Radke, T.C., Kaspar, T.B., Parkin, X., Jaynes, D.B., 1996. Soil property contributions to yield variation pattern. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Proceedings of the Third International Conference on Precision Agriculture. ASA, CSSA, SSSA, Inc, Madison, WI, pp. 417±424. Carlson, T.N., Ripley, D.A., 1997. On the relation between NDVI, fractional vegetation cover, and leaf area index. Remote Sens. Environ. 62, 241±252. Cora, J.E., Pierce, F.J., Basso, B., Ritchie, J.T., 1999. Simulation of within-®eld variability of corn yield with Ceres-Maize model. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Precision Agriculture. ASA-CSSA-SSSA, Madison, Wisconsin, USA, pp. 1309±1319. Gee, G.W., Bauder, J.W., 1986. Particle-size analysis. In: Methods of Soil and Plant Analysis I. American Society of Agronomy, Madison, Wisconsin, USA, pp. 383±411. Gamma Design Software, 1999. Gs+: geostatistics for the environmental scienc. Version 3.5. Gamma Design Software, Plainwell, MI, p. 44.
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
111
Hoogenboom, G.J., Jones, J.W., Wilkens, P.W., Batchelor, W.D., Bowen, W.T., Hunt, L.A. et al., 1994. Crop models. In: Tsuji, G.Y., Uehara, G., Balas, S. (Eds.), DSSAT v3., Vol. 2-2. University of Hawaii, Honolulu, Hawaii. Jackson, R.D., Huete, A.R., 1991. Interpreting vegetation indeces. Preventive Veterinary Medicine 11, 185±200. Jasinski, M.F., 1996. Estimation of subpixel vegetation density of natural regions using satellite multispectral imagery. IEEE Trans. Geosci. Remote Sens. 34, 804±813. Jones, J.W., Ritchie, J.T., 1991. Crop growth models. In: Homan, G.J., Howell, T.A., Soloman, K.H. (Eds.), Management of Farm Irrigation Systems. ASAE, St. Joseph, MI. Liu, Q., Huete, A., 1995. A feedback based modi®cation of the NDVI to minimize canopy background and atmosphere noise. IEEE Trans. Geosci. Remote Sens. 33, 457±465. Moran, M.S., Inoue, Y., Barnes, E.M., 1997. Opportunities and limitations for image-based remote sensing in precision crop management. Remote Sensing of the Environment 61, 319±346. Paz, J.O., Batchelor, W.D., Colvin, T.S., Logsdon, S.D., Kaspar, T.C., Karlen, D.L., 1997. Analysis of water stress eects causing spatial yield variability in soybeans. Trans. ASAE 41, 1527±1534. Paz, J.O., Batchelor, W.D., Babcock, B.A., Colvin, T.S., Logsdon, S.D., Kaspar, T.C., Karlen, D.L., 1999. Model-based technique to determine variable rate nitrogen for corn. Agric. Syst. 61, 69±75. Pierce, F.J., Nowak, P., 1999. Aspects of precision agriculture. Advances in Agronomy 67, 1±86. Pierce, F.J., Warncke, D.D., 2000. Soil and crop response to variable-rate limiting for two Michigan ®elds. Soil Sci. Soc. Am J. 64, 774±780. Pierce, F.J., Warncke, D.D., Everett, M.Wm, 1995. Yield and nutrient availability in glacial soils of Michigan. In: Robert, P.C. et al. (Eds.), Proc. of the Int. Conf. on Site Speci®c Management for Agricultural Systems, 2nd. Bloomington, WI. Pierce, F.J., Anderson, N.W., Colvin, T.S., Schueller, J.K., Humberg, D.S., Mc.Laughlin, N.B., 1997. Yield mapping. In: Pierce, F.J., Sadler, E.J. (Eds.), The State of Site Speci®c Management for Agriculture. ASA Misc. Publ., ASA, CSSA, and SSSA, Madison, WI, pp. 211±243. Price, J.C., 1992. Estimating vegetation amount from visible and near infrared re¯ectances. Remote Sens. Environ. 41, 29±34. Ritchie, J.T., 1987. Using crop models as a decision support system to reduce nitrate leaching. In: D'Itri, F.M., Wolfson, L.G. (Eds.), Rural Groundwater Contamination. Lewis Publishers, Chelsea, Michigan, pp. 179±192. Ritchie, J.T., 1991. Speci®cations of the ideal model for predicting crop yield. In: Muchow, R.C., Bellamy, J.A. (Eds.), Climatic Risk in Crop Production: Model and Management for the Semi-arid Tropic and Subtropic. Proc. Intl. Symp. S. Lucia, Brisbane, Queensland, Australia 2±6 July, 1990. CAB International, Wallingford, U.K., pp. 97±122. Ritchie, J.T., Crum, J., 1989. Converting soil survey characterization data into IBSNAT crop model input. In: Bouma, J., Berget, A.K. (Eds.), Land Qualities in Space and Time. Wageningen, the Netherlands, pp. 155±167. Ritchie, J.T., Godwin, D.C., Otter-Nacke, S., 1985. CERES-Wheat: a simulation model of wheat growth and development. Texas A&M University Press, College Station, TX. Ritchie, J.T., Gerakis, A., Suleiman, A., 1999. Simple model to estimate ®eld-measured soil water limits. Tran. ASAE 42, 1609±1614. Sadler, E.J., Russell, G., 1997. Modeling crop yield for site speci®c management. In: Pierce, F.J., Sadler, E.J. (Eds.), The State of Site Speci®c Management for Agriculture. ASA Misc. Publ., ASA, CSSA, and SSSA, Madison, WI, pp. 69±80. Sadler, E.J., Bauer, P.J., Busscher, W.J., 2000. Site-speci®c analysis of a droughted corn crop: I. growth and grain yield. Agron. J. 92, 395±402. Sudduth, K.A., Drummond, S.T., Birrell, S.J., Kitchen, N.R., 1996. Analysis of spatial factors in¯uencing crop yield. In: Robert, P.C., Rust, R.H., Larson, W.E. (Eds.), Proceedings of the Third International Conference on Precision Agriculture. ASA, CSSA, SSSA, Madison, WI, pp. 129±140. Verhagen, J., Bouma, J., 1997. Modeling soil variability. In: Pierce, F.J., Sadler, E.J. (Eds.), The State of
112
B. Basso et al. / Agricultural Systems 68 (2001) 97±112
Site Speci®c Management for Agriculture. ASA Misc. Publ., ASA, CSSA, and SSSA, Madison, WI, pp. 55±68. Verhagen, A., Booltink, H.W.G., Bouma, J., 1995. Site-speci®c management: balancing production and environmental requirements at farm scale. Agric. Syst. 49, 369±384. Wiegand, C.L., Richardson, A.J., Escobar, D.E., Gerbermann, A.H., 1991. Vegetation indeces in crop assessments. Remote Sens. Environ. 35, 105±119.