Active remote sensing and grain yield in irrigated maize
- First Online:
- Cite this article as:
- Inman, D., Khosla, R., Reich, R.M. et al. Precision Agric (2007) 8: 241. doi:10.1007/s11119-007-9043-z
- 245 Views
Advances in agricultural technology have led to the development of active remote sensing equipment that can potentially optimize N fertilizer inputs. The objective of this study was to evaluate a hand-held active remote sensing instrument to estimate yield potential in irrigated maize. This study was done over two consecutive years on two irrigated maize fields in eastern Colorado. At the six- to eight-leaf crop growth stage, the GreenSeeker™ active remote sensing unit was used to measure red and NIR reflectance of the crop canopy. Soil samples were taken before side-dressing from the plots at the time of sensing to determine nitrate concentration. Normalized difference vegetation index (NDVI) was calculated from the reflectance data and then divided by the number of days from planting to sensing, where growing degrees were greater than zero. An NDVI-ratio was calculated as the ratio of the reflectance of an area of interest to that of an N-rich portion of the field. Regression analysis was used to model grain yield. Grain yields ranged from 5 to 24 Mg ha−1. The coefficient of determination ranged from 0.10 to 0.76. The data for both fields in year 1 were modeled and cross-validated using data from both fields for year 2. The coefficient of determination of the best fitting model for year 1 was 0.54. The NDVI-ratio had a significant relationship with observed grain yield (r2 = 0.65). This study shows that the GreenSeeker™ active sensor has the potential to estimate grain yield in irrigated maize; however, improvements need to be made.
KeywordsGreenSeekerActive sensorNDVIGrain yieldMaize
Advances in technology have led to the development of active remote sensing systems that have their own energy source and are therefore not limited by the constraints that hinder other types of remote sensing (i.e., aerial and satellite imagery and aerial photographs). Active remote sensing systems are now available commercially. In principle, they can be mounted on a high-clearance tractor coupled with a fertilizer application boom and then used to vary the amount of fertilizer for a given area in ‘real-time’.
Reflection of light at the leaf-level depends primarily on pigments and internal structure of the leaf. Chlorophyll contained in the mesophyll of the leaf controls much of the visible light (400–720 nm) reflectance. Chlorophyll absorbs between 70 and 90% of all incident electromagnetic radiation in the blue and red (R) wavelength bands while reflecting light in the green (G) band (Campbell 2002). The amount of blue and red light absorbed by the leaf is directly proportional to the chlorophyll density of the leaf. On the other hand, reflectance of the near infrared (NIR) portion of the electromagnetic spectrum (720–1300 nm) is influenced primarily by the leaf’s mesophyll cells. The upper layers of the leaf are nearly transparent to NIR energy. Mesophyll tissue scatters and reflects as much as 60% of all incident NIR radiation. The degree to which NIR energy is reflected depends on the structure of the mesophyll cells and cavities between these cells (Campbell 2002).
Maize plants have 50–70% of plant N contained in the chloroplasts (Vleeshouwers and Jongschaap 2001). Chlorophyll concentration in the leaf and leaf N concentration are strongly related (r2 = 0.83) (Ercoli et al. 1993). The relative ‘greenness’ of a leaf can be used as a surrogate measure of leaf N concentration. Laboratory studies have shown that leaf N content is strongly related to green reflectance (Ercoli et al. 1993; Blackmer et al. 1994) and to the ratio of green to NIR reflectance (Schepers et al. 1996). The relationship between plant reflectance and N concentration has also been established in field-scale studies (Bausch and Duke 1996).
NDVI is a broadband index that is well correlated to leaf area index and green biomass (Peñuelas et al. 1994), and thus is sensitive to photosynthetic efficiency (Aparicio et al. 2002). Research has shown that NDVI is useful for estimating grain yield in certain crops. Raun et al. (2001) showed that expected yield determined from NDVI had a strong relation with actual grain yield in winter wheat (Triticum aestivum L.), (r2 = 0.83, p > 0.0001). Ma et al. (2001) reported that NDVI could be used to predict small and large yields in soybeans (Glycine max. L) reliably.
Active remote sensing has shown some potential for improving nitrogen-use efficiency (NUE) in winter wheat. Raun et al. (2001) found that NDVI calculated from mid-season (Feekes’ growth stages 4–6 in winter wheat, Large (1954)) reflectance measured with an active remote sensor could be used to estimate grain yield potential. They determined that using the number of growing-degree days from the day of planting to the day of sensing as a standardizing factor for NDVI resulted in regression models that could be applied across most of the nine site-years studied (Raun et al. 2001). However, these models could not account for conditions arising after sensing because the reflectances were measured during the middle of the growing season. Accounting for ‘post sensing’ crop stress is one of the limiting factors that hinders the development of a sensor-based N-rate algorithm. Nevertheless, NDVI standardized by growing-degree days accounted for over 50% of the variation in observed grain yield (Raun et al. 2001). In an attempt to develop a single regression model to estimate grain yield potential from NDVI, Raun et al. (2001) combined data from all sites even though NDVI varied considerably between sites.
Bausch and Diker (2001) showed that the NRI was a good predictor of plant N at the 9-leaf to 12-leaf crop growth stages. However, the effects of the soil background on reflectances had a negative effect on these relations (Bausch and Diker 2001). Raun et al. (2005) developed an NDVI response index (RINDVI) from the ratio of the mean NDVI of an N-rich area to the mean NDVI of an area receiving the field N rate. This response index was developed to account for variability in mineralized N supplied to the crop. Raun et al. (2002) found that when the RINDVI is combined with in-season grain yield estimates, NUE could be improved by more than 15% when compared to traditional farmer N application practices. In a large-scale study, Raun et al. (2005) found that the relation between in-season grain yield estimates and actual wheat grain yield (Mg ha−1) remained consistent over 30 site-years. They showed that the coefficients used in regression models did not differ significantly across time periods; suggesting that reliable and usable yield potential models can be developed with just 2 years of field data.
Despite the advancements made towards improving NUE for wheat using active sensing, there are few studies in the literature that address its use for irrigated maize. If active remote sensing is as promising for irrigated maize as it has been shown to be for wheat, it could have a tremendous impact on farm economics as well as the environment for maize producing regions.
The overall objective of this study was to evaluate the potential of the GreenSeeker™ hand-held active remote sensing instrument, which relies on red and NIR reflectances to estimate grain yield in irrigated maize. Specific objectives were: (a) to investigate the relationship between NDVI determined from active remote sensing and observed grain yield, (b) develop and evaluate regression models that estimate grain yield from active remote sensing data, and (c) investigate the relationship between reflectance measured using active remote sensing and soil nitrate concentration before side dressing with N.
Materials and methods
The study was done on two fields, A and B, over two consecutive growing seasons. The sites were in northeastern Colorado, USA, and were under a continuous maize cropping system. Field A was furrow irrigated and field B was center-pivot sprinkler irrigated. The soil of field A was mapped as Fort Collins Loam (fine-loamy, mixed, superactive, mesic Aridic Haplustalf). Field B was mapped as Albinas loam (fine-loamy, mixed, superactive, mesic Pachic Argiustoll), Ascalon fine sandy loam (fine-loamy, mixed, superactive, mesic, Aridic Argiustoll), and Haxtun loamy sand (fine-loamy, mixed, superactive, mesic Pachic Argiustoll) soil series.
Maize was planted at 77,000 plants ha−1 for field A and 84,000 plants ha−1 for field B. The spacing between rows was 76 cm for both fields. Field A was planted with Garst hybrid 8802 on day 126 (year 1) and day 122 (year 2). Field B was planted with Pioneer hybrid 34K77 on day 111 (year 1) and day 129 (year 2). Field A received a starter-band of 56 kg N ha−1 together with 8.6 cm of water at the time of planting and field B 34 kg N ha−1 together with 0.8 cm of water. No additional N applications were made during the crop growing season.
Soil samples for determining nitrate before side dressing were taken from each plot for both fields and years. Samples were taken from both the planting bed and furrow to a depth of 30 cm and mixed to give a composite sample. There were 48 composite pre-sidedress nitrate soil samples in total for both fields and years. Total soil NO3–N was determined by the method of Mulvaney (1996).
To evaluate the potential of GreenSeeker™ to estimate grain yield from reflectance measurements taken at the six- to eight-leaf crop growth stage, grain yield samples were taken at the crop’s physiological maturity (R6 crop growth stage) for grain yield analysis. One randomly located grain yield sample was taken from each plot for both sites and years; the sample comprised two 1-m long sections of a maize row. There were 48 grain yield samples in total (12 samples from two fields over 2 years). Grain yield mass was calculated as Mg ha−1 at 155 g kg−1 moisture.
Statistical analysis was done with SAS v8.0 (SAS Institute 2001) and SPLUS v.6.1 (Insightful corp., Troy NY, USA). Several regression models were investigated and compared for each site and year. Thenkabail et al. (2000) and Raun et al. (2005) showed that non-linear regression models are better for relating remote sensing-based vegetation indices to crop biophysical parameters, such as crop yield and biomass. Non-linear least-squares regression analysis, in Splus 6.0, was used to fit an exponential regression model relating observed grain yield to INSEY. An exponential model is often used in crop growth studies where the rate of crop growth at a given time is proportional to the amount of crop growth remaining (Neter et al. 1996). The choice of the exponential non-linear model was logical because INSEY is essentially an estimate of the rate of accumulated biomass (Stone et al. 1996). Other studies with the Greenseeker sensor have also shown that the relation between INSEY and grain yield is a non-linear exponential one (Raun et al. 2001, 2002, 2005).
Linear and non-linear least-squares regression analysis were used to regress grain yield on INSEY, grain yield on NDVI-Ratio, and pre-sidedress soil nitrate on NDVI-ratio. The significance of each regression model was tested using an F-test for lack of fit (Neter et al. 1996). The significance of the independent variables in each model was tested to determine if they were significantly different from zero with a two-tailed t-test. Regression parameters were considered to be significant at the p ≤ 0.05 level of significance. Independent variables that were significant at the p ≤ 0.05 level of significance were retained in the model. Residuals of the regression models were investigated by examining residual plots and testing the normality of residuals by the Shapiro-Wilk test (Neter et al. 1996).
Data from the GreenSeeker™ have been shown in other studies to estimate grain yield accurately from only 2 years of field data (Raun et al. 2005). Data from the two fields in the first year were used as the regression model building data set and data from the two fields for the second year were used as the regression model validation data set. The response variables (i.e., INSEY) of year 2 were incorporated into the regression models developed from year 1. For example, INSEY values for year 2 were incorporated into the regression models developed from year 1, and estimates of grain yield were made from this. These were then compared to the observed grain yield for year 2 for both fields. Both non-linear and linear models were compared based on their MSE, MSPE and bias results. A two-tailed t-test was used to determine if the bias in estimated grain yield was significantly different from zero.
Raun et al. (2005) suggested that the relation between INSEY and grain yield was consistent over 30 site-years, which has important implications for the utility and potential of GreenSeeker™. In addition to the stated objectives, we also evaluated the consistency of the relations between GreenSeeker™ data and grain yield across the sites used in this study, which had different soil types, planting densities, and hybrids. To assess the consistency of the relationship between INSEY and observed grain yield, the effect of the site and year (i.e., site-year) was included in the regression models. For this study, Fields A and B in the first year are referred to as site years I and II, respectively, and in the second year as site years III and IV, respectively. Since site-years are qualitative data, they were transformed into binary indicator variables (for more explanation of binary indicator variables see Neter et al. 1996) to analyze them quantitatively. The binary indicator variables were independent variables in the regression analysis (e.g., X1 = 1 if site-year 1 or III, X1 = 0 otherwise).
Results and discussion
Regression model, regression function, coefficient of determination (r2), mean squared error (MSE), mean squared prediction error (MSPE), estimate of bias (%), and p-value of the estimate of bias. All regression models are significant at p < 0.05
Ŷ (Mg ha−1) = −0.27 + 1599.3 (INSEYa)
Ŷ (Mg ha−1) = 84.5−17250 (INSEY) + 1016987.9 INSEY2
Ŷ (Mg ha−1) = 2.92 e(163.64 (INSEY))
Ŷ (Mg ha−1) = 28.4 + 19.47 ln (NDVI)
Ŷ (Mg ha−1) = 14.98−9.52(SYb) + 457.34 (INSEY)
Identifying areas of the field where there is sufficient and or insufficient plant-available N (i.e., areas of the field that will respond to fertilizer N) is integral to improving N fertilizer application strategies. Mean soil nitrate values were 0.8, 8.5, 4.8, and 8.2 ppm for field A and B in year 1 and field A and B in year 2, respectively. As expected, observed grain yield was moderately affected by soil nitrate (r2 = 0.51, p < 0.05). Similar results have been reported in the literature (Magdoff et al. 1984; Bundy and Andraski 1995; and Rozas et al. 2000). According to Spellman et al. (1996), nitrate values before side dressing of <13 ppm in the top 30 cm indicate the need for additional N. Based on this critical level of nitrate, it was below the threshold for both fields in both years, and some additional N fertilizer would have been required to optimize grain yields. However, it has been suggested that the soil nitrate concentration before side dressing tends to over-estimate N responsiveness (Meisinger et al. 1992; Bundy and Andraski 1993; Heckman et al. 1995). The NDVI-ratio used in this study compares the NDVI of an area of interest to that of an area that has sufficient N, and it should indicate if an area has the potential to respond to additional N. For example, it has been reported that an NRI of 0.95 or less indicates N deficiency and the area should have additional N fertilizer (Bausch and Duke 1996; Bausch and Diker 2001). Overall, the NDVI-ratio increased with increasing nitrate values; although the relation was statistically significant it was weak (r2 = 0.31, p < 0.05). Magdoff (1991) reported that N fertilizer recommendations based on soil nitrate (measured before side dressing) can be unreliable when used on fields that have large amounts of rainfall and permeable soil because of the potential for nitrate to leach from the top 30 cm of the soil. The fields used in this study had coarse-textured soil and were irrigated (furrow irrigation for field A and center-pivot sprinkler for field B), and therefore the use of pre-sidedress soil nitrate to predict fertilizer need requirements was probably not ideal.
Conclusions and future directions
This study shows that the GreenSeeker™ active sensor has the potential to estimate grain yield in irrigated maize; however, there is much room for improvement. We observed a linear relationship between grain yield and INSEY, which was significantly different between fields and years. Variability between sites and growing seasons limits the precision with which data from the Greenseeker™ sensor can be used to model potential yield. Other researchers have used this tool to improve nitrogen use efficiency in wheat with great success. Our results suggest that there is potential to do the same in irrigated maize. More research is needed, however, before this is a reality.
Areas that need additional study in future include, but are not limited to the following: (i) collecting data from more locations and years to determine if the relation between INSEY and maize yield is stable, (ii) determine the critical level for the NDVI-ratio used in this study, (iii) determine the critical growth stage at which use of the sensor would be most beneficial, (iv) does mounting the GreenSeeker™ to farm equipment change any of the fundamental relationships and or does it introduce interference from the soil background, and (v) how many sensors need to be mounted on each farm equipment for different crops. Before active sensors such as GreenSeeker™ become practical on farm equipment, the benefits (economic, environmental, and nitrogen use-efficiency) must outweigh the effects of the costs and skills needed to incorporate active sensors into farming practices. While more needs to be done, this and other studies are steps in the right direction.