DETECTION OF NITROGEN STRESS IN CORN USING DIGITAL AERIAL IMAGING Sreekala GopalaPillai Lei Tian and John Beal ABSTRACT High-resolution color infrared (CIR) aerial images were used for detecting in-field spatial patterns, especially due to nitrogen stress in a corn field. Digital images were processed and classified into groups using unsupervised learning method and analyzed using a GIS package. The CIR images indicated soil type in the earlier stages of crop growth and nitrogen stress during the later stages of crop growth. The canopy reflectance was well correlated to the applied nitrogen and the yield from 75days after sowing. The yield could be predicted fairly accurately (R 2 = 0.96) from the canopy reflectance of red light. Incorporation of temporal variations along with spatial variability for yield prediction would enhance the accuracy of prediction. Aerial remote sensing proves to be a promising tool for obtaining spatial and temporal in-field variability in the crop field for site-specific management and yield prediction. INTRODUCTION Though not a new idea, precision farming and associated technology has become the thrust area in the recent past. Precision farming concept is centered on the fact that the productivity of soil varies spatially and temporally within a field, depending on the environmental, soil and operational conditions. Hence the soil productivity and the final profit from the land can be increased tremendously by implementing precision farming through variable rate technology (VRT). Through VRT, management practices like chemical application and tillage can be done site-specifically to suit the local requirements within each field. The biggest difficulty in implementing VRT is the lack of a reliable and consistent method of obtaining spatial and temporal variability data for the field (Sawyer, 1994). Aircraft remote sensing is important in precision crop management 1
(PCM), for monitoring seasonally variable soil and crop conditions for time-specific and time critical crop management (Moran et al., 1997). While satellite remote sensing has the limitations of time and resolution, aircraft remote sensing has high resolution and time flexibility. Aerial photography was reported as a promising tool for assessing the variability in the crop due to spatially varying nitrogen status in the field (Blackmer et al., 1995; Filella et al., 1995; Tomer et al., 1997). Reflectance indices were used by all the above scientists, as an indicator of nitrogen stress in the field. Reflectance based crop indices are accepted as a very useful means for evaluating crop stress (Jackson et al., 1986). Reflectance at 550 nm, 680 nm (Chappelle, 1992) and near infrared at 800nm (Adcock et al., 1990) were found to be highly indicative of the Chl A (Chlorophyll A) content and hence the nitrogen availability to the plant. Single band reflectance indices like R550 (reflectance at 550nm), R675 (reflectance at 675nm), red edge position (λ RE = maximum slope of vegetative reflectance spectra between red and near infra red region), pigment simple ratio (PSR = R430/R680) and normalized pigment chlorophyll ratio (NPCI) were used for classifying crops based on the nutrient status (Filella et al., 1995). Reflectance based vegetative indices like normalized difference vegetative index (NDVI) and ratio of NIR to red reflectance are reported to be better indicators of nitrogen status than single band reflectance indices (Ma et al., 1996). Penuelas et al (1994) used physiological reflectance index (PSI) and NPCI to identify physiological changes in nitrogen limited sunflower. Aerial photographs, both color (RGB) and color infrared (CIR) were used to predict nitrogen uptake and yield (Blackmer et al., 1995; Tomer et al., 1997) as well as for yield map interpretation (Bausch and Duke, 1996). All the above scientists used film 2
cameras to capture the image and then scanned the image into the computer for further processing and analysis. This could have resulted in significant loss in both the spatial and spectral resolution and signal to noise ratio. Low spatial resolution, poor signal to noise ratio and poor and variable lighting conditions are some of the bottlenecks in applying this technology. There have been rigorous efforts to overcome these bottlenecks. It would be interesting to see the scope of remote sensing in site-specific crop management (SSCM) with an airborne high-resolution digital camera. It remedies some of the problems due to low resolution and improves the signal to noise ratio as it do not use photographs and scanners. The general objective of this study was to evaluate the digital aerial imaging technique for precision farming in-field variability sensing. In this study we concentrated on one such in-field variability factors namely soil type and nitrogen stress. The digital aerial imaging technique was tested for delineating the spatial and temporal in-field variability in nitrogen uptake in a cornfield and to predict the yield based on the variability data obtained from CIR images. MATERIALS AND METHODS Color Infrared (CIR) images indicate crop conditions by capturing the reflected near infra red (NIR), red and green radiation from the crop canopy. The NIR reflected from a plant varies significantly at different stages of the plant growth as well as with different problems that might cause stress in the crop. CIR images of the experimental cornfields in Champaign and Decatur were taken during the period of May to September 1997, using a high-resolution DCS 420 CIR camera (Eastman Kodak, Rochester, New York) integrated with a GPS receiver (TRIMBLE Ensign XL), from an aircraft (1996 3
Mooney M20-E) for aerial application. The aerial imaging system set up used in this study for near-real time multi-function biological target remote sensing was described in detail by Tian et al (1997). The experiment field in Champaign was used mainly for soil type analysis and comparison. The nitrogen test field was a 75 A plot with half of the area under nitrogen treatment. Nitrogen levels of 0, 60, 120, 180 and 240lb/A were applied in the field with two replications at 180 level and 4 replications at 60 level. At the end of the season both soil nitrate and stalk nitrate were tested. The yield data was collected using a yield monitor (MicroTrak - Graintrak). The nitrogen test field was planted on April 19, 1997 and harvested on Sept 25, 1997, 160days after sowing (DAS). The crop was harvested a little early due to the damage caused by wind. The CIR images were taken on 55, 75, 99, 125, 141 and 147DAS, the last two set of images being taken after the wind damage. Stalk nitrate was tested for all the nitrogen plots on the day of harvest. The digital images stored during the flight in a PCMCIA card in DCS formats (original compressed kodak format) were acquired into TIF files using Adobe Photoshop 3.0.4 software. The camera was used with a color infrared filter (650BP300) to get CIR images (500-810nm) and a color filter (550BP300) to get visible light image. The acquired images were 3 channel IRRG (Infrared-Red-Green) with CIR filter and 3 channel RGB (Red-Green-Blue) with visible light filter, 12 bits per channel, 1524X1012 pixels and 4.4 MB size. The resolution of the camera varies with elevation of the aircraft. At an elevation of 450m the image resolution is 0.2m/pixel and at 220m elevation the resolution is 0.1m/pixel. Only color infrared images were used in this study. The downloaded images were geo-referenced to its geographic co-ordinates using the boundary map collected with a vehicle mounted GPS system (Ashtech BR2G 4
differential) and plotted into ArcView shape file using AGRISS software. All the images were geo-referenced using a Geographic Transformer V3 (BlueMarble Geographics) to a lower field resolution or 4ft per pixel, for comparability between the images taken at different elevations, as well as to analyze them using Geographic Information Systems (GIS). The geo-referenced, 3-chanel digital images taken early in the season were analyzed initially to compare reflectance data from the field with the soil types. But later in the season after the crop canopy took over the soil, the images were used to compare the reflectance pattern from the nitrogen plots for plant stress. Chromaticity diagrams and histograms were used for the initial analysis of the pixel distribution in the images. The distribution canopy reflectance was used to find out the number of clusters an image may have. All the images were clustered (unsupervised classification) using ISODATA procedure in MultiSpec software (Purdue Univ). The clustered images were compared with soil map, yield map and application map using GIS for spatial patterns in the crop field. The average canopy reflectance values for each nitrogen plot obtained from the aerial images were correlated to the applied nitrogen and stalk nitrate to find the potential of aerial remote sensing to detect nitrogen stress. The average canopy reflectance is also correlated to the yield to find out the best growth period for predicting the yield based on the reflectance indices. The best-correlated image was used to develop a yield model. RESULTS AND DISCUSSION The red (R) and green (G) component of the three-channel CIR image was used as reflectance indices for soil type and nitrogen stress studies. Part of the infra red (IR) 5
component in the images of the nitrogen test field was saturated and hence the vegetative indices like NDVI were not used for this set of images. Soil type patterns Results of the initial analysis of the images are shown here. The histogram of a sample area of the image is checked for the number of different reflectance distributions present in the image. The histogram of the research farm is shown in Figure 1. Both red and green channels show four distinguishable distributions whereas infrared channel shows 5 distinct distributions in the canopy reflectance. These distributions indicate the spatial variability factors in the field 17DAS. The plant cover over soil was insignificantly less and the patterns, hence, was assumed to be due to the soil variability. Soil type was considered as one of the prime source of this variability. In Figure 2 the canopy reflectance values from areas under the five major soil type in the field were plotted with different legends, on color co-ordinates of IR, R and G. The reflectance values formed five separate groups in the color space. This strongly indicates that the soil type contributed the patterns observed in that particular image. The grouping is more distinguishable in the scatter plot on IR, R, and G co-ordinate system than in the normalized two-dimensional chromaticity diagram (Figure 3) where all the three-channel reflectance values are normalized by the intensity component (division by IR+R+G). The different groups overlap each other in the normalized diagram and there is no linear boundary between the different groups. 6
Frequency (%) 100 80 60 40 20 Infrared Red Green 0 17 37 57 77 97 117 137 157 177 197 Gray level Figure 1 Histogram of research farm image, 17DAS. The infrared histogram shows 5 distinct distributions whereas the other two channels indicate 4 groups. 1. Cluster1 2. Cluster2 3. Cluster3 4. Cluster4 X Column 3 5 Cluster5 Components Y Column 4 Z Column 5 y z x Figure 2 Pixel distribution on IR, R, G co-ordinates for research farm, 17DAS. IT shows five clusters corresponding to the five soil types 7
Normalized infrared, ir 0.5 0.4 0.3 0.2 0.1 cluster1 cluster2 cluster3 cluster4 cluster6 0 0 0.1 0.2 0.3 0.4 0.5 Normalized red,r Figure 3 Chromaticity diagram (using normalized IR and R) of sample area from research farm image 17DAS. With the information derived from the histogram and scatter plot on color coordinates, all the geo-referenced images were processed and classified using the unsupervised clustering algorithm, ISODATA. The CIR image of research farm, taken 17DAS is shown in Figure 4. This image was processed (clustering and classification) and overlaid with the soil map and the boundary map. The soil map and clustered image showed the same kind of patterns (Figure 5) for research farm showing very little error in the spatial overlay. This suggests that aerial remote sensing of bare soil would be a very reliable way of mapping soil types. 8
Figure 4 An example CIR image of research farm 17DAS indicating the soiltypes. Shades of green are soil and shades of red are plants. (a) (b) Figure 5 (a) Soil map overlaid with clustered image and (b) map overlay indicating error between soil type and clustered pattern For nitrogen test farm, a detailed soil map was not available. Hence the county soil map was used as the reference map for comparing the soil type and the soil patterns observed in the image. But as shown in Figure 6, the cluster pattern observed in the image was different though it resembled to the pattern in the soil map. This suggests that the county soil map is not very accurate. Choosing the soil sampling grids according to 9
the soil patterns in the processed image would improve the accuracy of soil type mapping compared to the conventional grid sampling method. SoilType, Image: Beal.tif SoilType,C l6-13,image: B eal2.cluster.tif (a) CIR image overlaid with soil map (b) Clustered image with cluster boundary overlaid with soil map Figure 6 Comparison of county soil map with the soil pattern obtained from aerial CIR images. Both images indicate the error in county soil map Nitrogen Stress Detection Images acquired early in the season for the nitrogen tests didn t show any indication of nitrogen stress. This is mainly due to the low coverage of the plant canopy over the soil. The no-nitrogen plots started showing up the stress patterns in the images from 75DAS onwards. As indicated by the later images, stress developed in the 60lb/A nitrogen plots at a much later time. The scatter plot of pixels in color co-ordinates, for different nitrogen plots in the experimental farm 125DAS is shown in Figure 7. The three groups of no-nitrogen, 60lb/A nitrogen and well nourished (greater than 120lb/A 10
nitrogen) formed three clear groups though their boundary is not very distinct as in the soil types. It is hard to differentiate between the high nitrogen rates in the image. N0 N60 N>=120 G IR R Figure 7 Distribution of pixels on IR, R, G co-ordinates for an experimental nitrogen field. Different colors indicates three nitrogen levels The three-channel canopy reflectance was correlated to the nitrogen applied in the soil as well as stalk nitrate for all the images. Figure 8 shows the variation of correlation coefficient between applied nitrogen and the R and G canopy reflectance. The IR channel was saturated in this particular set of images and hence it was not used for correlation purpose. Both R and G channels are negatively correlated to the nitrogen level and hence R/G is positively correlated. In the earlier stages of crop growth, there was no significant correlation between the applied nitrogen and reflection indices. One reason for such a low correlation is that the nitrogen in the soil was enough for the germinating plants in the beginning but later on they developed stress due to lack of nitrogen. It could also be due to the low canopy coverage in the earlier stages of plant growth. Segmenting out the soil 11
from the image may help to find the nitrogen stress at an earlier stage. A maximum correlation coefficient of -0.8 was observed on 125DAS with R and G channels, but it dropped down again to non-significant levels as the crop matured. This drop is due the substantial damage in the crop due to wind before the last two sets of images were taken. The images after the wind showed wind damage patterns more prominently than the nitrogen stress. After the crop matures, it starts drying and the chlorophyll content reduces at this time. So, the canopy reflectance changes as the chlorophyll content reduces and it would not indicate the nitrogen availability as accurately as compared to an actively growing plant. Correlation coefficient 0.8 0.4 0-0.4-0.8 55 75 99 125 141 147 R G R/G -1.2 Growth period (DAS) Figure 8 Variation of correlation coefficient between applied nitrogen and reflection indices at different growth stages of corn 12
0.8 Correlation coefficient 0.4 0-0.4 55 75 99 125 141 147 R G R/G -0.8 Growth period (DAS) Figure 9 Variation of correlation coefficient between stalk nitrate and reflection indices at different growth stages of corn The correlation between stalk nitrate and reflectance values is shown in Figure 9. This correlation was not very significant throughout the growth period. The stalk nitrate was correlated to neither applied nitrogen (r = 0.46) nor yield (r = 0.40). The low correlation is because the fallen crop had to meet its nitrogen requirements from the stalk nitrogen since it couldn t get any nitrogen from the soil after the wind damage. This must have depleted the stalk nitrogen, depending on the extent and pattern of damage caused by the wind. The stalk nitrate of a healthy plant may be well correlated to the applied nitrogen, yield and reflectance indices. Yield Analysis Figure 10 shows the yield map overlaid with application map and soil map. For yield analysis, the yield map was overlaid with processed (ISODATA clustered) images on 75, 99, 125, 141, and 147DAS. A spatial analysis was done to obtain the average yield under each cluster for the nitrogen test field. This average yield per cluster is correlated to the average reflectance for the cluster, to find the pattern that defines the yield the best. 13
Figure 11 shows the variation of the correlation coefficient between cluster centers and yield. A maximum correlation of 0.94 was obtained between yield and R-reflectance with image on 125DAS. Generally the correlation was high (above 0.8) from 75 to 141DAS. This high correlation indicates that the yield varies according to the spatial patterns observed in the CIR image and it can be predicted from the canopy reflectance indices. # -# yield map(low-high) soil map application map Figure 10 Yield map of Decatur farm overlaid with application map and county soil map. 14
0.4 Correlation coefficient 0-0.4-0.8 55 75 99 125 141 R G -1.2 Growth period (DAS) Figure 11 Variation of correlation coefficient between average yield under cluster patterns and the cluster center at different growth stages of corn 200 175 150 125 100 75 75 100 125 150 175 200 Actual yield (B/A) Figure 12 Yield predicted from canopy reflectance 125DAS plotted against actual yield. Slope = 0.96, Intercept = 6.36 The CIR image taken 125DAS was used to fit a model for predicting the yield from the canopy reflectance values. Though a linear fit on R reflectance (yield = 631.5-2.5R) was good with an R 2 value of 0.89, multiple regression yielded a polynomial of second degree on R reflectance (yield = 0.07 R 2 +22.7 R 1762) with an R 2 value of 0.96. 15
Figure 12 shows the predicted values plotted against the actual yield. It shows a good prediction, with a slope of 0.96 and intercept of 6.36. This model used only the spatial infield variation data from one image for predicting the yield. Incorporation of the temporal variation of in-field spatial patterns along with the spatial variation would improve the prediction of yield. CONCLUSIONS The histograms, scatter plots of IR, R and G reflectance and the processed results of the CIR images showed clear grouping of pixels, indicating the spatial variability within the image. These spatial patterns resembled to the predominant factor (soil or crop) of the time. A spatial analysis of the images of the field (mainly soil) early in the crop season showed that the patterns represent the soil type. Processed aerial images of bare soil indicates soil type and it could be used as a guide for soil sampling in the field for soil analysis instead of relying on conventional grid sampling method. The CIR aerial images could delineate the nitrogen stress areas in the field whereas differentiating between the nitrogen levels of well-nourished areas was difficult. The three channel canopy reflectance values were significantly correlated to the applied nitrogen. There was no good correlation between the stalk nitrate taken during harvest and the canopy reflectance during the entire growing season. This was mainly due to the wind damage in the field prior to harvest. The canopy reflectance was well correlated to the yield from 75DAS to 141DAS, with a maximum correlation on 125DAS. The spatial patterns observed through the CIR images reflect factors that have affected the yield. The predicted yield from canopy red reflectance was a good estimate of the actual yield. Incorporation of the temporal in-field 16
variability may result in better results on yield prediction. Aerial imaging is a promising method for collecting in-field variability data essential for site-specific crop management and yield prediction. ACKNOWLEDGEMENT The authors thank the Illinois Council of Food and Agricultural Research for supporting this project. REFERENCES Adcock, T. E., F. W. Nutter, and P. A. Banks. 1990. Measuring herbicide injury to soybeans (Glycine max) using a radiometer. Weed Sci. 38:625-627. Bausch, W. C and H. R. Duke. 1996. Remote sensing of plant nitrogen status in corn. Transactions of ASAE. 39(5):1869-1875. Blackmer, T. M., J. S. Schepers, and G. E. Meyer. 1995. Remote sensing to detect nitrogen deficiency in corn. Proceedings of Site-Specific Management for Agricultural Systems. Second International Conference: ASA, CSSA, SSSA, Madison WI. P.502-512. Chappelle, E. W., M. S. Kim and J. E. McMurrtrey III. 1992. Ratio analysis of reflectance spectra (RARS): an algorithm for remote estimation of the concentrations of chlorophyll a, chlorophyll b, and carotenoids in soybean leaves. Remote Sens. Environ. 39(3):239-247. Filella, I., L. Serrano, J. Serra and J. Penuelas. 1995. Evaluating wheat nitrogen status with canopy reflectance indices and discriminant analysis. Crop Sci. 35:1400-1405. Jackson, R. D., P. J. Pinter, R. J. Reginato and S. B. Idso. 1986. Detection and evaluation of plant stresses for crop management decisions. IEEE Transactions on Geoscience and Remote Sensing. GE-24(1):99-106. Ma, B. L., M. J. Morrison and L. M. Dwyer. 1996. Canopy light reflectance and field greenness to assess nitrogen fertilization and yield of maize. Agron J. 88:915-920. 17
Moran, M. S., Y. Inoue and E. M. Barnes. 1997. Opportunities and limitations for image based remote sensing in precision crop management. Remote Sens. Environ. 61:319-346. Penuelas, J., J. A. Gamon, A. L. Fredeen, J. Merino and C. B. Field. 1994. Remote Sens. Environ. 48:135-146. Sawyer, J. E. 1994. Concepts of variable rate technology with considerations for fertilizer application. J. of Production Agric. 7:195-201. Tian, L., R. Hornbaker, and R. Schmidt. 1997. Aerial field sensing and mapping for precision farming. Paper No. AA97-001. Presented in NAAA/ASAE Joint Technical Session at Las Vegas, Nevada. Tomer, M. D., J. L. Anderson and J. A. Lamb. 1997. Assessing corn yield and nitrogen uptake variability with digitized aerial infrared photographs. Photogrammetric Engineering & Remote Sensing. 63(3):299-306. 18