Influence of Hybrid, Population, and Nitrogen Rate on Spectral Prediction of Corn Grain Yield

With the escalation in environmental concern and cost of production, researchers have focused on investigating more efficient means of increasing grain yield while reducing fertilizer use. This study was conducted to evaluate the use of normalized difference vegetation index (NDVI) and derived coefficient of variation (CV) to predict grain yield in corn (Zea mays L.) as effected by hybrid, plant population, and fertilizer N rate. A quadratic-plateau model existed between NDVI and plant population, resulting in a critical NDVI, ranging from 0.74 to 0.78 in red NDVI (RNDVI) and 0.65 to 0.66 in green NDVI (GNDVI). Conversely, the CV relationship with plant population indicates that CV can recognize increased plant biomass at levels well above the NDVI plateau and therefore can be utilized with NDVI to determine plant biomass for yield prediction. The most effective growth stage to predict grain yield was V8, presumably because the highest variability in NDVI and CV occurred at this stage. Hybrid maturity did not effect grain yield prediction at V8, but reproductive growth stage yield prediction may require hybrid maturity categorization. The NDVI and CV determined using either the red or green bands were equally effective for use in predicting corn grain yield from early season sensor measurements.

As environmental concerns continue to escalate and agricultural production becomes more scrutinized, new fertilizer application practices will continue to be researched with the goal of increasing fertilizer use efficiency. Currently, the Environmental Protection Agency (EPA) is reporting that watersheds in all 48 states of the continental U.S. tested for nitrate nitrogen (NO₃^-N) have groundwater contamination levels above the maximum contaminant level (MCL), of which Oklahoma is ranked 14^th(EPA, 1999a). Production of cereal grains has largely been held responsible for this groundwater contamination, in particular corn (Zea mays L.) production, where high nitrogen (N) rates have been applied in high yielding environments. Excessive N applications to cereal grain crops continue to pollute the environment, increasing human health risk and costing farmers needless additional expense along with negative publicity. This exemplifies the need for continued research to improve fertilizer use efficiency.

Raun and Johnson (1999) reported worldwide nitrogen use efficiency (NUE) estimates to be approximately 33%, with developing countries estimated at 29% and developed countries at 42%. As a result, N fertilizer losses were valued at about $15.9 billion dollars annually, which as of August of 2001 has increased to $20 billion dollars annually with the price of N fertilizer nearly doubling due to the shortages of natural gas (Raun et al., 2002). There are several paths in which N can be lost in the N cycle, dependent heavily on the crop and environmental conditions surrounding it.

As split applications of N became more popular, different methods have been used to measure early season fertilizer use as a means of deciding how much more N fertilizer would be needed to meet the predetermined grain yield goal. Several researchers looked at in-season soil test of NO₃^-N at a depth of 30 cm, known as presidedress NO₃^-N test (PSNT). Research from Vermont showed that use of a PSNT reduced N rates without reducing grain yields in corn by setting an N application limit with no N sidedress applications needed when the PSNT is above 25 mg kg^-1 NO₃^-N (Magdoff et al., 1984; Durieux et al., 1995). While the PSNT worked well for the northeastern USA, some adjustments were made to the PSNT for the semi-arid regions under irrigation, particularly reducing the no N application limit to about 15 mg kg^-1 NO₃^-N (Spellman et al., 1996). The PSNT test was further refined in corn production by Bundy and Andraski (1995) when they reported that classifying soils into two categories medium and high yield potential determined by: depth of root zone, water holding capacity, and length of growing season, when the test values were considered in the N responsive range. While the PSNT improved NUE, it could not produce these results consistently, even when other soil properties were evaluated in tandem with PSNT.

Chlorophyll meters (SPAD meters) have been successfully used to determine in-season N status, since chlorophyll content has been highly correlated with leaf N concentration (Wolfe et al., 1988; Schepers et al., 1992). With the chlorophyll meters, researchers developed an N Sufficiency index [(as-needed treatment/ well-fertilized treatment) * 100] from which recommendations were made for in-season N fertilizer applications when the index values fell below 95% (Blackmer and Schepers, 1995; Varvel et al., 1997). Varvel et al. (1997) reported that maximum grain yields in corn were attained when early season sufficiency indexes ranged between 90 and 100% up to the V8 growth stage, but if the sufficiency index fell below 90% at V8, maximum yields were not realized due to early season N deficiency resulting in lost yield potential. Peterson et al. (1993) indicated that variation in chlorophyll meter measurements can range up to 15% from plant to plant, requiring considerable measurements in order to maintain a representative average for the field at each sampling date. Nevertheless, the chlorophyll meter was not a viable tool for guiding N side-dress decisions for corn in the Southern Coastal Plain, undoubtedly because there was no variable-rate application determination in the sufficiency index used, either apply or don’t apply (Gascho and Lee, 2002). Another drawback of the chlorophyll meter is that by reading one leaf at a time, plant biomass cannot be determined as with the remote sensor.

Using a photodiode-based remote sensor measuring canopy radiance in the red (671nm) and near-infrared (NIR, 780nm) spectral bands, Stone et al. (1996) developed a plant N spectral index (PNSI) for correcting in-season wheat N deficiencies. This index, the absolute value of the inverse of the normalized difference vegetation index (NDVI), saved between 32 and 57 kg N ha ^–1compared to fixed N rates. Bausch and Duke (1996) developed an N Reflectance Index (NRI), using a canopy reflectance ratio of NIR (760-900nm)/ green (520-600nm) for the low N area to the NIR/ green for a well N-fertilized area, a very similar procedure to the sufficiency index differing only in how the measurement is taken. Recent work has shown that in-season normalized difference vegetation index {NDVI = [(NIR_ref/NIR_inc) – (Red_ref/Red_inc)]/[(NIR_ref/NIR_inc) + (Red_ref/Red_inc)]} highly correlated with final grain N uptake (Lukina et al., 2001; Raun et al., 2002). Therefore, an in-season response index (RI) from NDVI, referred to as RI_NDVI (Highest mean NDVI N treatment/ Mean NDVI check treatment), was evaluated and determined to be a viable method for measuring the potential response to additional N (Mullen et al., 2003).

GopalaPillai and Tian (1999) reported that by using linear regression models on normalized intensity (NI) of a high-resolution color infrared (CIR) field image, they were able to predict grain yield with 55 to 91% accuracy, depending on the field and growing season. However, their work conceded that better correlation with yield was obtained when using CIR images taken after pollination, than those from earlier in the season, similar to the findings of work from Bartholome (1988) evaluating better correlation to grain yield with accumulated NDVI after booting in millet (Panicum miliaceum L.) and grain sorghum (Sorghum bicolor L.). With the goal of improving correlation of NDVI to grain yield, Rasmussen (1998) integrated the product of multitemporal NDVI with photosynthetically active radiation, but was not successful since no single regression line was valid for consecutive years. Other yield prediction models have been developed for corn, although they do not use in-season plant health evaluations as part of the equation. The crop environment resource synthesis (CERES)-Maize model simulates the major physiological processes, soil water movement processes, and N transformation and transport processes involved in plant growth and has been studied a great deal and has been determined to simulate grain N uptake (Pang et al., 1997). The artificial neural network (ANN) model was successful in predicting grain yield, designed to evaluate the following three aspects of 15 input factors: (1) yield trends with temperature, rainfall, soil texture, and soil pH, (2) interaction between N application and late July rainfall, and (3) optimization of the input factors with a genetic algorithm (Liu et al., 2001). However, both models use extensive evaluation of multiple factors to predict final grain yield by full season simulation, but were not accurate for mid-season grain prediction. As a function of increasing NUE research, this study was conducted to evaluate the potential of using NDVI to determine N response and predict grain yield in corn. The effects of corn hybrid, plant population, and fertilizer N rate on NDVI were also evaluated to establish what adaptations might be necessary to use NDVI over a wide range of field conditions.

Two experimental sites were established in the spring of 2002, one near Stillwater, OK at the Lake Carl Blackwell Agronomy Research Farm (Pulaski fine sandy loam, course-loamy, mixed, nonacid, thermic Typic Ustifluvent), and one near Haskell, OK at the Eastern Oklahoma Research Station (Taloka silt loam fine, mixed, thermic Mollic Albaqiustoll). Initial soil test results are reported in Table 1. The experiment employed a factorial arrangement within a randomized complete block design with three replications. Individual plots measured 3.0 x 9.1 m with 76.2 cm row spacing in conventional tillage. Ammonium nitrate (34-0-0) was preplant surface-broadcasted by hand at rates of 0, 56, and 112 kg N ha^-1.

Three Bacillus thuringiensis (bt) gene enhanced corn hybrids identified by their maturity date (105-day, 109-day, and 113-day) were planted at both sites in 2002, but three different hybrids without bt gene enhancement (104-day, 107-day, 111-day) were planted at both sites in 2003 (planting, fertilizer, and harvest dates are reported on Table 2). Two different seeding rates were evaluated at both sites: at the Haskell site seeding rates of 44,460 (low) and 66,690 seeds ha^-1 (high) in 2002 and 49,400 (low) and 71,630 seeds ha^-1 (high) in 2003 and at the Lake Carl Blackwell (LCB) site, seeding rates of 35,568 (low) and 51,870 seeds ha^-1 (high).

Each corn plot was sensed with a GreenSeeker™ Hand Held optical reflectance sensor (Ntech Industries, Ukiah, CA), measuring RNDVI {RNDVI = [(NIR_ref/NIR_inc) – (Red_ref/Red_inc)] / [(NIR_ref/NIR_inc) + (Red_ref/Red_inc)]} and coefficient of variation (CV = standard deviation/mean) of the RNDVI at different vegetative and reproductive growth stages at both sites each year (sensing dates and growth stages with description presented in Table 3) with the sensor nadir to the ground and approximately 70 cm above the crop canopy. Corn grain was harvested with a Massey Ferguson 8XP experimental combine, removing 2 rows from the center of each plot. A Harvest Master yield-monitoring computer installed on the combine was used to record grain weight and moisture levels. Grain yield from each plot was determined by adjusting grain weight to 15.5% moisture and a grain sub-sample was taken for total N analysis.

In the spring of 2004 the experiment was reconfigured with the relocation of the Haskell trial to an adjacent field at the same location with the same soil description. Furthermore, plant populations were increased from two populations to four (37050, 51870, 66690, and 81510 seeds ha^-1) and corn hybrids were reduced from three to two Bacillus thuringiensis (bt) gene enhanced corn hybrids (99-day and 113-day). Initial soil test results and planting, fertilizer, and harvest dates for the relocated Haskell site are also reported in Tables 1 & 2. Individual plot size was reduced to 3.0 x 6.1 m with the purpose of sustaining manageable labor requirements for the increased treatment size. In addition, ammonium nitrate (34-0-0) application rates were increased to 84 and 168 kg N ha^-1 to assure N would not be a limiting factor in high N application treatments.

With the availability of a green NDVI sensor, both GNDVI {GNDVI = [(NIR_ref/NIR_inc) – (Green_ref/Green_inc)] / [(NIR_ref/NIR_inc) + (Green_ref/Green_inc)]} and RNDVI were measured at different vegetative and reproductive growth stages at both sites (sensing dates and growth stages are presented in Table 3) with GreenSeeker™ Hand Held optical reflectance sensors. In addition, the CV derived from both GNDVI (GCV) and RNDVI (RCV) was evaluated. The center two rows of each corn plot were sensed separately with the sensor nadir to the ground and approximately 70 cm above the crop canopy. Corn grain was harvested (picked and shucked) by hand from the center two rows of each plot separately and ear weights were recorded for each row. Four random ears from each row were collectively weighed, dried in a forced air oven at 66^oC, and weighed again to determine moisture levels. Following the measurement of dry weights, the four ears were shelled by hand using a Root-Healey Manufacturing Company (Plymouth, OH) hand-crank corn sheller and the grain weight was taken to determine an average cob weight for each row. Finally, grain yield from each row was calculated by adjusting grain weight to 15.5% moisture and a grain sub-sample was taken for total N analysis.

Grain samples were dried in a forced air oven at 66^oC, ground to pass a 140 mesh sieve (100 um), and analyzed for total N content using a Carlo-Erba NA 1500 automated dry combustion analyzer (Schepers et al., 1989). Nitrogen use efficiency was determined using the difference method: dividing the difference between the grain N uptake of the N treatment and the grain N uptake of the check (0 N rate) by the N rate of the N treatment. Analyses of variance and single degree of freedom contrasts were performed using SAS (SAS, 2002). Linear and non-linear regression models were used to determine the relationships present between grain yield and NDVI as well as between both NDVI and CV and the treatment variables.

Significantly higher RNDVI and lower RCV values occurred in the higher plant population at both locations for both the vegetative and reproductive sensor readings (data not shown). Although plant biomass increased considerably between the V10 and R1 sensor readings, RNDVI values decreased extensively due to tassel development at Haskell. The lighter color of the tassel decreased red light absorbance in the crop canopy. At both sites RCV declined between the first and second sensor readings, indicating that an inverse relationship occurs between RNDVI and RCV, as reflectance decreased from increased biomass absorbance, less variability was detected.

The high plant population was significantly greater in RNDVI for all three hybrids at the first three sensor readings (V6, V7, and V8) and at the V10 growth stage for the 104-day and 111-day hybrids. However, no differences were found within the plant populations of any hybrid at the V16 growth stage. Although significant treatment effects were visually present at V16, the effects could not be measured effectively with RNDVI. Decreased RNDVI values were observed in the high plant population of the 104-day hybrid at the R3 and R4 reproductive growth stages as well as the R4 sensing of the 107-day hybrid. However, the 111-day hybrid RNDVI increased in the higher plant population at the R1 and R2 growth stages. While the 111-day hybrid may have obtained significantly higher RNDVI values than either of the other two hybrids during the reproductive growth stages, this occurred because of the maturity differences between the hybrids (as the hybrid matures RNDVI decreased due to hastened lower leaf senescence) since the 104-day and 107-day hybrids produced significantly higher grain yields. The low plant population was significantly higher in RCV than the high population for all hybrids at the V6, V7, and V8 growth stages. The low plant population in the 104-day hybrid was higher in RCV than the high population at V16 as well, but had a lower RCV at R4. The low plant population was higher in RCV than the high population for both the 107-day and 111-day hybrids from R1 to R3, but the high plant population was higher at R4 in the 107-day hybrid.

The high plant population at the LCB site increased the RNDVI values and decreased the RCV values of all three hybrids at all sensor (V6-R2) readings. At LCB, the decease in RNDVI did not occur at tassel development, but this was essentially due to the lower plant populations used. The greatest vegetative difference for both RNDVI and RCV between the plant populations occurred around V8. The V8 sensor data revealed that a quadratic-plateau relationship existed between measured plant population and RNDVI for all three hybrids when combining both sites and excluding measurements from the 0 N treatments (Figure 1). The 0N treatments decreased RNDVI due to inadequate N, resulting in high variability in RNDVI that overshadowed the variability associated with plant population. The power function model best described the relationship between RCV and plant population when combining both sites for all three hybrids (Figure 2).

A positive linear response to plant population in both GNDVI and RNDVI occurred for both hybrids at all (V5-R5) sensor readings over both sites and negative linear responses to plant population in GCV and RCV were seen at all sensor readings. The quadratic response to plant population was variable between locations, hybrids, and NDVI spectrums, however quadratic plateau models best described the relationship between both GNDVI and RNDVI and plant population when the hybrids were separately evaluated at the V8 growth stage (Figures 3 & 4). Power function models best described the relationship between both GCV and RCV and Plant population for both hybrids (Figures 5 & 6). The RNDVI relationship to plant population was greater for both hybrids than the GNDVI relationship, although the two NDVI spectrums had similar curves in both hybrids as was the case with RCV compared to GCV.

As shown by the quadratic plateau models (Figures 1, 3, & 4), the RNDVI plateau occurred in a narrow range between 0.74 and 0.78, however the plant population at which the plateau developed ranged greatly (between 49,000 & 62,000 plants ha^-1) reducing the accuracy of the combined model (Figure 7). This indicated that at an RNDVI measurement of 0.76, increased biomass is no longer distinguishable, while the plant population at which that occurs varies with hybrid and environmental conditions. Similarly, the RCV relationship with plant population was reduced in the combined model (Figure 8), however the power function confirms that RCV can estimate plant biomass at RNDVI levels above the plateau.

At Haskell a positive linear response to N was seen for RNDVI in all three hybrids at the V10 reading, but no N response was observed for RCV. While visual differences were observed in the trial at the R1 reading, no significant trends were determined from either RNDVI or RCV measurements and there was no linear response to N in grain yield for any of the three hybrids at either plant population. Positive linear responses to N were measured in RNDVI and RCV at LCB for all hybrids at both readings, which was consistent with the grain yield data.

A positive linear RNDVI response to N occurred at Haskell in both plant populations of all three hybrids at the V6, V7, and V8 sensor readings. As mentioned beforehand canopy closure occurred at V16, at which time it was not possible to distinguish observable N deficiencies (Figure 9). Negative linear RNDVI responses to N were evaluated in the reproductive growth stages, particularly after the R1 growth stage for both plant populations of all three hybrids at Haskell. Negative linear RCV responses to N were observed at Haskell for V6, V7, V8, and R1 for all hybrids, but positive linear RCV responses to N were observed at R3 and R4 for all hybrids (Figure 10). At tassel development RNDVI values declined, however, the RNDVI values of the lower fertility treatments declined less since tassel development was associated with plant health. Grain yield showed a linear response to N in the 104-day hybrid, but not in either of the other two hybrids. The lack of significant N response in the 107-day and 111-day hybrids signifies late-season stress reduced grain fill and consequently the effects of N deficiency since the earlier maturing 104-day hybrid was significantly higher in grain yield. At LCB a positive linear RNDVI response to N occurred in the low plant population of the 107-day hybrid for the R1 reading, no other RNDVI responses to N were observed for either plant population of any hybrid and no N responses were found in RCV. Grain yield data was consistent with RNDVI and RCV measurements; no significant N response was found for any hybrid and/or plant population at LCB.

At Haskell positive linear GNDVI and RNDVI responses to N were observed for both hybrids in all plant populations during mid (V6-V9) and late (R4-R5) season sensor readings (Figure 11). Negative linear GCV responses to N occurred for the 99-day hybrid in the 66,690 and 81,510 plant populations at V8 and in the 66,690 at V9. In the 113-day hybrid negative linear GCV responses were observed in the 37,050 plant population between V8 and V11 and in the 66,690 plant population at V8. RCV was more sensitive to N response at Haskell with negative linear responses for the 99-day hybrid in the 37,050 plant population at V5, V8, V9, and V11 and in the 51,870 plant population at V5. Negative linear RCV responses in the 113-day hybrid occurred in the 37,050 plant population at all sensor readings except V5 and in all four of the plant populations at V8 (Figure 12). Grain yield showed a positive linear N response in the 37,050 plant population of both hybrids and a quadratic N response in all of the higher plant populations of both hybrids.

Positive linear GNDVI responses to N observed at LCB were inconsistent within the 113-day hybrid with sporadic responses occurring in the 37,050 and 66,690 plant populations and consistent responses occurring in the 51,870 and 81,510 plant populations between V7 and R4 growth stages. Positive linear GNDVI responses to N in the 99-day hybrid at LCB were not seen for the 37,050 population the entire season or the other plant populations until the reproductive growth stages. Positive linear RNDVI responses to N were limited to the 51,870 plant population of both hybrids and the 81,510 of the 113-day hybrid between V7 and R4 growth stages. Negative linear GCV responses to N were randomly observed in the reproductive stages of the 113-day hybrid and no response in the 99-day hybrid. No responses to N in RCV were observed for either hybrid as well. Grain yield data indicated no N response in the 99-day hybrid and a positive linear response in the 81,510 plant population of the 113-day hybrid only. In general, the red band (RNDVI and RCV) was more sensitive to N response than the green (GNDVI and GCV), particularly in CV (Figures 11 & 12). Furthermore, the greatest variation between the high N and 0 N treatments in both NDVI (GNDVI & RNDVI) and CV (GCV & RCV) occurred approximately at the V8 growth stage, consistent with the RNDVI and RCV results from 2003. Suggesting that the most accurate measurement of biomass N response happens at V8, since high green reflectance and red absorbance limits NDVI accuracy later in the vegetative stages.

The linear regression model was evaluated between grain yield and RNDVI measurements and between grain yield and RCV in 2002. Although the sensor measurements may have some inaccuracy due to maintaining improper height above the crop canopy (i.e. holding the sensor too close to the canopy), some very pronounced relationships were found. Comparisons between grain yield and both RNDVI and RCV at early growth stages (V7 at LCB, V10 at Haskell) resulted in poor relationships at both sites (Tables 4 & 5), but comparisons made at the later reading (R1 growth stage) showed a very good relationship between grain yield and both RNDVI and RCV. The R1 sensor readings may have been more effective since tassel development was affected by plant health and therefore narrowing the sensing field of view by holding the sensor too close to the crop canopy would not greatly affect the RNDVI measurement from a plant health prospective. At earlier growth stages before tassel development, RNDVI measurements would need to measure plant biomass which potentially could not be done effectively if the sensor is held too close to the crop canopy, resulting in less variability (lower RCV) between treatments and inflated RNDVI readings. Therefore mid-season RNDVI can be utilized very effectively to predict grain yield.

Crop Year 2003

Linear regression analysis revealed strong relationships between grain yield and RNDVI at early growth stages (V6, V7, V8, and V10) in 2003 (Table 4). Similar results were observed from linear regression between grain yield and RCV (Table 5). These data indicate that early season grain yield prediction is achievable and therefore sidedress N application based on grain yield prediction is practical. Although lower than the V6 and V7 relationships with grain yield, the V8 growth stage as previously mentioned showed the greatest variation between plant populations and N rates, revealing that V8 growth stage sensor readings would give the greatest indication of vegetative biomass growth whether or not it translates to grain yield (Figures 13 & 14). The R2 growth stage comparison between grain yield and RNDVI also showed a well-defined relationship supporting the 2002 results that late season yield prediction is possible though not practical for side-dress N application (Table 4). However, the R2 model showed that each hybrid has a separate linear relationship with grain yield corresponding to 2002 results and the model was improved significantly to that of the V8 data when the hybrids were fitted with separate lines (data not shown). Separating out the hybrids in the early season model did not improve the relationship with grain yield, confirming that while significant differences in RNDVI occurred between the hybrids at V8 these differences existed in grain yield as well. Mid-season grain yield prediction can be achieved in the reproductive stages as supported by the 2002 data, and in earlier vegetative growth stages, particularly at V8 when the highest RNDVI variability occurred between treatments.

Crop Year 2004

In 2004, the V8 data showed the highest positive linear relationship between both GNDVI and RNDVI with grain yield (Table 4). However, the NDVI (both GNDVI & RNDVI) and CV (both GCV & RCV) relationships with grain yield were considerably higher in the 99-day hybrid compared to the 113-day hybrid. Therefore developing separate yield prediction lines for these two hybrids would have been necessary to maintain accurate grain yield prediction. Little rationale can be given as to why grain yield prediction using both NDVI and CV was improved for the 99-day hybrid. Plant height differences between hybrids were small at V8, as early maturing hybrids did not typically separate from the longer season hybrids in this study (over 3 years) until the reproductive stages. The most plausible explanation could be that the 99-day hybrid was not as well suited for the dry-land environment as anticipated and was under substantial stress throughout the growing season that limited yield potential early. Therefore, the 99-day hybrid was potentially less sensitive to the post-sensing (late-season) environment. While the red spectrum (RNDVI & RCV) had a distinct advantage over the green (GNDVI & GCV) in relation with plant population and N response, there were no advantages observed between the spectrums and grain yield (Tables 4 & 5). Poor relationships resulted from grain yield and the NDVI measurements taken in the reproductive growth stages (R1-R5), possibly due to a violent thunderstorm with damaging hail and strong winds that occurred between the V9 and V10 growth stages (June 2) that caused severe plant damage from torn leaves and stunted plants due to lodging as proven by the lower RNDVI values at V11-VT.

The critical population at which the RNDVI plateau occurred ranged between 49,000 and 62,000 plants ha ^–1, however RNDVI plateau occurred in a narrow range between 0.74 and 0.78. This indicated that increased biomass is no longer distinguishable when RNDVI reaches 0.76, while the plant population at which that occurs varies with hybrid and environmental conditions. The GNDVI resulted in similar results in 2004, although the RNDVI had a distinct advantage over GNDVI. Conversely, the RCV relationship with plant population indicated that RCV can recognize increased plant biomass at levels well above the RNDVI plateau and therefore can be utilized with RNDVI to determine plant biomass for yield prediction.

The NDVI data from the V8 growth stage predicted grain yield accurately in 2003 and 2004, presumably because the highest variability in NDVI and CV occurred at the V8 growth stage both years. Later vegetative growth stages may actually contain more plant variability than at V8 and could have stronger relationships with grain yield. However canopy closure occurs shortly after V8 (V10 to V12) and vegetative stage NDVI data collected thereafter cannot accurately account for plant variability. Well-defined relationships between NDVI and grain yield also occurred in the reproductive growth stages in 2002 and 2003, but at different growth stages each year: R1 in 2001, R2 in 2003. Although late-season yield prediction is not useful for N management and limited due to temporal variability, the potential is there for other uses.

Separating the hybrids improved these reproductive relationships with grain yield all three years, but only improved the V8 relationship with grain yield in 2004. Hybrid maturity did not effect grain yield prediction at V8, but reproductive growth stage yield prediction will require hybrid maturity categorization. Finally, comparisons made between the GNDVI and RNDVI relationships with grain yield in 2004 showed no significant differences over three locations.

Bartholome, E. 1988. Radiometric measurements and crop yield forecasting; Some observations over millet and sorghum experimental plots in Mali. Int. J. Remote Sens. 9:1539-1552.

Bausch, W.C. and H.R. Duke. 1996. Remote sensing of plant nitrogen status in corn. Trans. ASAE 39(5):1869-1875.

Blackmer, T.M. and J.S. Schepers. 1995. Use of a chlorophyll meter to monitor nitrogen status and schedule fertigation for corn. J. Prod. Agric. 8(1):56-60.

Bundy, L.G. and T.W. Andraski. 1995. Soil yield potential effects on performance of soil nitrate tests. J. Prod. Agric. 8:561-568.

Durieux, R.P., H.J. Brown, E.J. Stewart, J.Q. Zhao, W.E. Jokela, and F.R. Magdoff. 1995. Implications of nitrogen management strategies for nitrate leaching potential: Roles of nitrogen source and fertilizer recommendations system. Agron. J. 87:884-887.

Gascho, G.J. and R.D. Lee. 2002. Determining side-dress nitrogen requirements of corn following broiler litter in the Southern Coastal Plain. J. Plant Nutr. 11:2361-2371.

GopalaPillai, S. and L. Tian. 1999. In-field variability detection and spatial yield modeling for corn using digital aerial imaging. Trans. ASAE 42(6):1911-1920.

Liu, J., C.E. Goering, and L. Tian. 2001. A neural network for setting target corn yields. Trans. ASAE 44(3):705-713.

Lukina, E.V., K.W. Freeman, K.J. Wynn, W.E. Thomason, R.W. Mullen, A.R. Klatt, G.V. Johnson, R.L. Elliott, M.L. Stone, J.B. Solie, and W.R. Raun. 2001. Nitrogen fertilization optimization algorithm based on in-season estimates of yield and plant nitrogen uptake. J. Plant Nutr. 24:885-898.

Magdoff, F.R., D. Ross, and J. Amadon. 1984. A soil test for nitrogen availability to corn. Soil Sci. Soc. Am. J. 48:1301-1304.

Menezes, R.S.C., G.J. Gascho, W.W. Hanna, M.L. Cabrera, and J.E. Hook. 1997. Subsoil nitrate uptake by grain pearl millet. Agron. J. 89:189-194.

Mullen, R.W., K.W. Freeman, W.R. Raun, G.V. Johnson, M.L. Stone, and J.B. Solie. 2003. Identifying an in-season response index and the potential to increase wheat yield with nitrogen. Agron. J. 95:347-351.

Pang, X.P., J. Letey, and L. Wu. 1997. Yield and nitrogen uptake prediction by CERES-Maize model under semiarid conditions. Soil Sci. Soc. Am. J. 61:254-256.

Peterson, T.A., T.M. Blackmer, D.D. Francis, and J.S. Schepers. 1993. Using a chlorophyll meter to improve N management. NebGuide G93-1171-A. Cooperative Extension, Institute of Agriculture and Natural Resources, Univ. of Nebraska-Lincoln.

Rasmussen, M.S. 1998. Developing simple, operational, consistent NDVI-vegetation models by applying environmental and climatic information. Part II: Crop yield assessment. Int. J. Remote Sensing 19(1):119-139.

Raun, W.R., and G.V. Johnson. 1999. Improving nitrogen use efficiency for cereal production. Agron. J. 91:357-363.

Raun, W.R., J.B. Solie, G.V. Johnson, M.L. Stone, R.W. Mullen, K.W. Freeman, W.E. Thomason, and E.V. Lukina. 2002. Improving nitrogen use efficiency in cereal grain production with optical sensing and variable rate application. Agron. J. 94:815-820.

SAS Institute. 2002. The SAS system for windows version 8.0. SAS Institute Inc., Cary, NC.

Schepers, J.S., D.D. Francis, and M.T. Thompson. 1989. Simultaneous determination of total C, total N, and ¹⁵N on soil and plant material. Commun. Soil Sci. Plant Anal. 20:949-959.

Schepers, J.S., D.D. Francis, M. Vigil, and F.E. Below. 1992. Comparison of corn leaf nitrogen concentration and chlorophyll meter readings. Commun. Soil Sci. Plant Anal. 23(17-20):2173-2187.

Spellman, D.E., A. Rongni, D.G. Westfall, R.M. Waskom, and P.N. Soltanpour. 1996. Pre-sidedress nitrate soil testing to manage nitrogen fertility in irrigated corn in a semi-arid environment. Commun. Soil Sci. Plant Anal. 27:561-574.

Stone, M.L., J.B. Solie, W.R. Raun, R.W. Whitney, S.L. Taylor, and J.D. Ringer.1996. Use of spectral radiance for correcting in-season fertilizer nitrogen deficiencies in winter wheat. Trans. ASAE. 39(5):1623-1631.

U.S. Environmental Protection Agency (EPA). 1999a. National Primary Drinking Water Regulations Technical Fact Sheets. Washington, D.C.: Office of Water, Office of Ground Water and Drinking Water. http://www.epa.gov/OGWDW/hfacts.html. March

Varvel, G.E., J.S. Schepers, and D.D. Francis. 1997. Ability for in-season correction of nitrogen deficiency in corn using chlorophyll meters. Soil Sci. Soc. Am. J. 61:1233-1239.

water and nitrogen effects on senescence of maize: II. Photosynthetic decline and longevity of individual leaves. Agron. J. 80:865-870

Figure 1. Relationship between RNDVI and plant population of three hybrids at the V8 growth stage over two locations with 0 N treatments removed fitted to a quadratic-plateau model, 2003.

Figure 2. Relationship between RCV and plant population of three hybrids at the V8 growth stage over two locations with 0 N treatments removed, 2003.

Figure 3. Relationship between plant population and both GNDVI and RNDVI at V8 with 0 N treatments removed for the 99-day hybrid over two locations fitted to a quadratic-plateau model, 2004.

Figure 4. Relationship between plant population and both GNDVI and RNDVI at V8 with 0 N treatments removed for the 113-day hybrid over two locations fitted to a quadratic-plateau model, 2004.

Figure 5. Relationship between plant population and both GCV and RCV at V8 with 0 N treatments removed for the 99-day hybrid over two locations, 2004.

Figure 6. Relationship between plant population and both GCV and RCV at V8 with 0 N treatments removed for the 113-day hybrid over two locations, 2004.

Figure 7. Relationship between plant population and RNDVI at V8 with 0 N treatments removed for five hybrids over four site-years fitted to a quadratic-plateau model.

Figure 8. Relationship between plant population and RCV at V8 with 0 N treatments removed for five hybrids over four site-years.

Figure 9. Effect of N rate on RNDVI over time in the 104-day hybrid at the Haskell site, 2003.

Figure 10. Effect of N rate on RCV over time in the 104-day hybrid at the Haskell site, 2003.

Figure 11. Effect of N rate on both GNDVI and RNDVI over time in the 113-day hybrid at the Haskell site, 2004.

Figure 12. Effect of N rate on both GCV and RCV over time in the 113-day hybrid at the Haskell site, 2004.

Figure 13. Relationship between grain yield and RNDVI of the V8 growth stage over two locations, 2003. The 95% lower and upper individual sample confidence intervals are represented by the bottom and top lines, respectively.

Figure 14. Relationship between grain yield and RCV of the V8 growth stage over two locations, 2003. The 95% lower and upper individual sample confidence intervals are represented by the bottom and top lines, respectively.

Table 1. Initial surface (0-15cm) soil test results prior to experiment initiation at Haskell and Lake Carl Blackwell (LCB), OK, 2002-04.

Sample	NH₄-N	NO₃-N	P	K	pH
Sample	------------------- mg kg^-1 -------------------				pH
Haskell 2002-03	7.65	2.66	48	136	6.18
LCB 2002-04	14.28	1.63	23	239	6.13
Haskell 2004	NA	20.00	112	218	6.70

NH₄-N and NO₃-N – 2 M KCL extract; P and K – Mehlich-3 extraction; pH – 1:1 soil:deionized water; NA- not available Haskell site relocated in 2004.

Table 2. Planting, fertilizer, and harvest dates (dd-mm-yy) at Haskell and Lake Carl Blackwell (LCB), OK, 2002-04.

Location	Crop Year	Planting	Fertilizer Application	Grain Harvest
Haskell	2002	18-4-2002	16-4-2002	11-9-2002
LCB	2002	23-4-2002	23-4-2002	28-8-2002
Haskell	2003	03-4-2003	03-4-2003	20-8-2003
LCB	2003	01-4-2003	08-4-2003	07-8-2003
Haskell	2004	01-4-2004	01-4-2004	31-8-2004
LCB	2004	03-4-2004	03-4-2004	28-8-2004

Table 3. Sensing dates (dd-mm-yy format) by growth stage at Haskell and Lake Carl

Vegetative growth stages (V#) determined by number of collared leaves. Depending on hybrid and environmental conditions plants produce between 11 to 20 collared leaves (V11-V20), in 2004 only 12 collared leaves were observed before tasseling. --- No data available.

Table 4. Relationship between grain yield and NDVI by growth stage and crop year fitted to a linear regression model, 2002-04.

GNDVI = green normalized difference vegetation index [(NIR_ref/NIR_inc) – (Green_ref/Green_inc)]/[(NIR_ref/NIR_inc) + (Green_ref/Green_inc)]

NDVI = red normalized difference vegetation index [(NIRref/NIRinc) – (Redref/Redinc)]/[(NIRref/NIRinc) + (Redref/Redinc)]

Table 5. Relationship between grain yield and CV by growth stage and crop year fitted to a linear regression model, 2002-04.