Precision farming has been reinvented as a

Transcription

1 IN-FIELD VARIABILITY DETECTION AND SPATIAL YIELD MODELING FOR CORN USING DIGITAL AERIAL IMAGING S. GopalaPillai, L. Tian ABSTRACT. High-resolution color infrared (CIR) images acquired with an airborne digital camera were used to detect infield spatial variability in soil type, crop nutrient stress, and to analyze spatial variability in yield. Images were processed using an unsupervised learning (clustering) method. The clustered images were geo-referenced, and spatially analyzed using a GIS package. The image patterns in a processed image of bare soil matched well with soil type map with 76% accuracy. The CIR images of a cornfield indicated nitrogen stress areas from 75 days after planting (DAP). The CIR reflectance was better correlated to the yield after pollination of corn compared to the early images. The spatial variation in yield was linearly correlated to the spatial variation of individual reflectance bands (NIR, R, and G) as well as normalized intensity (NI) of CIR image. Spatial yield models on uncalibrated reflectance bands of image could predict 76 to 98% of yield variation in each field. A linear regression model on NI developed from one field image predicted yield with an accuracy of 55 to 91% in different fields and seasons. Digital aerial imaging proves to be a promising tool for obtaining spatial in-field variability in the crop field for site-specific management and yield prediction. Keywords. Color infrared image, In-field spatial variability, GIS, Nutrient stress, Soil type, Yield model. Precision farming has been reinvented as a revolutionary concept that could enhance crop productivity and decrease pollution and farming expenditure at the same time by reducing farm inputs to ideal levels. The precision farming concept is based on the fact that crop productivity varies spatially and temporally within a field, depending on environmental and soil conditions and operational activities. Hence, managing the field site-specifically in order to compensate the spatial variability in the field will increase the crop productivity along with the economic profitability. The technique of applying inputs in a spatially varying fashion is called variable rate technology (VRT). Using VRT, management practices like chemical application and tillage can be done site-specifically to match the local requirements within each field. Nevertheless, the biggest difficulty in implementing VRT is the lack of a reliable and consistent method of obtaining spatial and temporal variability data for a field (Sawyer, 1994). Remote sensing has currently been considered and experimented as a potential tool for collecting spatial and temporal variability data within a crop field for precision farming. Research on both satellite and aircraft remote sensing has been done for developing methods and techniques for in-field spatial variability data collection and their application in precision farming. These remote sensing techniques sense the reflected radiation from the plant Article was submitted for publication in October 1998; reviewed and approved for publication by the Information & Electrical Technologies Division of ASAE in October Presented as ASAE Paper No The authors are Sreekala GopalaPillai, ASAE Student Member, Graduate Student, Lei Tian, ASAE Member Engineer, Assistant Professor, Agricultural Engineering Dept., University of Illinois at Urbana-Champaign, Illinois. Corresponding author: Lei Tian, University of Illinois, 1304 W. Pennsylvania Ave., Urbana, IL 61801, voice: (217) , fax: (217) , lei-tian@uiuc.edu. canopy and/or soil surface. By analyzing the spectral reflectance from the crop field, important information on crop condition can be obtained. The near infrared (NIR) and visible spectrum of light reflected from a plant varies significantly at different stages of the plant growth as well as with crop stresses. Healthy plants tend to reflect more NIR radiation than stressed plants. During the vegetative growth period, the chlorophyll content increases and hence the absorption of visible light increases and its reflectance decreases. The relative decrease in green (G) reflectance is much less compared to red (R) and blue reflectance (Knipling, 1970), resulting in the green peak. Both higher nitrogen availability and chlorophyll contents in the plant results in lower green peak as compared to lower nitrogen availability and chlorophyll (Chappelle et al., 1992). The chlorophyll content and the green reflectance stabilize when the crop reaches maturity and decrease when the crop goes through senescence. The NIR reflectance and transmittance increases with vegetative growth and increased biomass due to the multiple reflection in the internal mesophyll of the plant (Bauer, 1985). The potential of remote sensing to identify soil properties and problems that affect crops were recognized by the scientific community as early as in 1930s (Curran, 1985). Spectral reflectance from any soil is determined by various physical and chemical properties of the soil such as moisture, organic matter, texture, surface roughness, and iron oxide (Bauer, 1985). Several studies on soil reflectance were done with the purpose of eliminating the effect of soil background from the reflectance data of vegetation canopies (Huete, 1988; Qi et al., 1994). There are no reported and standard methods of developing soil maps from aerial images of soil. In conventional method, the soil is visually assessed, sampled based on the visual assessment, and tested for developing soil type maps. Most of the reported studies on nitrogen status of crop used reflectance based crop indices for extracting crop Transactions of the ASAE VOL. 42(6): American Society of Agricultural Engineers / 99 /

2 stress information (Jackson et al., 1986). Reflectance values at 550 nm and 680 nm (Chappelle et al., 1992), and near infrared reflectance at 800 nm (Adcock et al., 1990) were found to be highly indicative of the Chlorophyll-A content and hence the nitrogen availability to the plant. Vegetative indices like pigment simple ratio (PSR) and normalized pigment chlorophyll ratios (NPCI) were used for classifying crops based on the nutrient status (Filella et al., 1995). Vegetative indices like normalized difference vegetative index (NDVI) and simple ratio index (SRI is same as PSR) were better indicators of nitrogen status than single band reflectance indices (Ma et al., 1996). Penuelas et al. (1994) used physiological reflectance index (PRI) and NPCI to identify physiological changes in nitrogen limited sunflower. Early and fast detection of crop stress is very important for taking appropriate remedial measures before the damage becomes irreversible. Field images taken early in this study suggested that it was possible to delineate the soil types and nitrogen stress at a high resolution using the digital CIR images (Gopalapillai et al., 1998). Yield prediction ahead of harvest is very crucial for both farmers and grain dealers for predicting the market and deciding on post-harvest handling of grains. The difficulty in predicting yield is that numerous factors related to weather, crop, soil, and operational activities affect the final yield. Yield prediction based on reflectance or vegetative indices may be able to overcome this difficulty. Bausch and Duke (1996) used aerial images for yield map interpretation. In their study on aerial remote sensing and yield variability, Senay et al. (1998) grouped the yield map into different classes and compared the mean yield of each class to the mean spectral reflectance for that class. They found that yield is linearly related to mean spectral reflectance (r = 0.99) and digital elevation model (r = 0.92). They also reported a high correlation between yield and near infrared bands compared to visible bands. A yield model based on the spatially varying spectral reflectance values would therefore be a good technique for analyzing and predicting yield for crops like corn. For precision agriculture applications, satellite remote sensing systems have several drawbacks when compared to aircraft remote sensing systems. Satellite remote sensing systems have inflexible or fixed imaging time and low spatial resolution for precision farming purposes. In the case of an aerial system, the timing and frequency of flight can be adjusted depending on the weather conditions or other requirements. Aircraft remote sensing is a reliable tool for monitoring seasonally variable soil and crop conditions for time-specific and time critical crop management (Moran et al., 1997). 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). Aerial photographs have also been used for yield map interpretation (Bausch and Duke, 1996). Although aircraft remote sensing has advantages like high resolution and time flexibility compared to satellite systems, several of the existing aerial imaging systems have drawbacks like relatively low spectral resolution, low accuracy, and problems associated with film development and image scanning. Film development and digitization by scanning may result in significant loss in spectral resolution resulting in poor signal to noise ratio in the images. The new near-real-time digital aerial imaging system (Tian et al., 1997) we used in this research capitalizes on several advantages including high spatial and spectral resolution, time flexibility, and relatively high signal accuracy. The general objective of this study was to evaluate the high-resolution, color-infrared digital aerial imaging technique for sensing in-field variability in soil type, nitrogen stress and yield, for precision farming applications. The specific objectives of this study are given below. 1. Study the potential of high-resolution digital aerial imaging technique for identifying soil types and to compare the spatial patterns in a soil image to a conventional soil map. 2. Test the high-resolution CIR imaging system for identifying nitrogen stress under different field conditions and nitrogen applications. 3. Analyze the spatial in-field variability in yield, and to develop a model to predict the spatial yield based on the CIR reflectance data obtained through remote sensing of the field. MATERIALS AND METHODS The experiment was conducted in three farms, the Agricultural Engineering Research farm (AGDE) located in Champaign County, a farmer s farm (FF) in Macon County, and Williams research farm (WF) in McLean county, all in central Illinois. The FF field data was collected in 1997 and the rest of the data was collected in All the fields have corn and soybean on rotation in alternate years. At the time of data collection, the fields were under corn. GROUND TRUTH PREPARATION For soil studies, soil data from AGDE farm was used. A detailed soil map of the AGDE farm developed in 1961 was used for comparing the soil pattern observed in CIR images. The soil map was prepared by first measuring the slope of the field, followed by visual evaluation of the soil and soil sampling wherever a different soil type was suspected. This soil map had a few more details but very similar as compared to a county soil map. The nitrogen rate experiment was conducted in 1997 corn in FF farm and in 1998 corn in AGDE farm. The nitrogen application experiments in both fields were unbalanced randomized designs. Nitrogen levels of 0, 67, 134, 202, and 269 kg/ha were applied to the nitrogen field in FF farm. There were five replications for 67 kg/ha, two replications of 202 kg/ha, and single replication for all other levels in the FF field. The nitrogen plots consisted of m strips for the 0 and 67 kg/ha applications, and m strips for the rest of the applications. The east part of the AGDE field was divided into 12 plots of m strips; with one zero nitrogen plot, two 56 kg/ha plot, and one 112 kg/ha plot. The remaining eight plots were applied with 168 kg/ha of nitrogen. The west part of the m AGDE field was under herbicide application for weed research and was applied with 168 kg/ha of nitrogen. In the herbicide plot, there were different applications of control, broadleaf, grass, and total (grass and broadleaf) herbicide 1912 TRANSACTIONS OF THE ASAE

3 in m strips. The nitrogen application was a completely randomized unbalanced design and separate from herbicide application. In 1997 the FF field had crop damages due to high wind one month before harvest. In the 1998 season, there was extensive rain and flooding that might have caused leaching of nitrogen, especially in low altitude areas. In 1998 the low-altitude areas of the WF field had to be replanted due to extensive damages caused by the rain. Stalk nitrate was tested in all nitrogen plots in both FF and AGDE fields. Three samples were collected from each plot and send to a laboratory (Midwest Laboratories, Inc, Omaha, Nebr.) for testing. One set of spadmeter readings was collected for the nitrogen plots in AGDE farm on 24 July The yield data was collected using MicroTrak Graintrak yield monitor (Micro-Trak Systems, Inc., Eagle Lake, Minn.) and the boundary data was collected using an Ashtec BR2G GPS/Beacon receiver (Ashtech, Sunnyvale, Calif.) in FF field. Image data and on the go yield data collected from FF field in 1997, and AGDE and WF fields in 1998 were used for analyzing and modeling the spatial variability in yield. For yield analysis, the yield data from the whole AGDE field, including the nitrogen and herbicide plots, were used. AERIAL IMAGING SYSTEM The aerial imaging system consists of a high-resolution digital camera, a GPS receiver, a portable computer, and a camera mount with three degrees-of-freedom. The DCS 420 CIR camera (Eastman Kodak, Rochester, N.Y.) has a single CCD sensor array (KAF1600) that has a spatial resolution of pixels, and spectral response in the range of 400 to 1000 nm. The camera was used with a color infrared filter (650BP300, Eastman Kodak, Rochester, N.Y.) that blocked the blue spectrum. The spectral range of CIR image was 500 to 810 nm. The CCD sensor array within the camera divided this spectrum into three bands of NIR (710 to 810 nm), Red (600 to 710 nm), and Green (500 to 600 nm). These three bands were then combined to generate a digital CIR image. The image bands were uncalibrated digital numbers in the range of 0 to 255 and they are referred to as reflectance or spectral bands in this article. These digital numbers were not the actual reflectance but the apparent radiance that fall on the sensor, which was the sum of the intrinsic radiance of the surface object (soil and crop) subjected to atmospheric transmittance, and the atmospheric path radiances (Bauer, 1985). The camera was integrated with a TRIMBLE Ensign XL GPS receiver (Trimble Navigation, Sunnyvale, Calif.) to record the GPS locations of the camera. The camera was controlled with a portable computer with a Pentium II 133 MHz processor. The camera was connected to the computer through a SCSI cable. The imaging system was mounted on a Cessna 205 fix wing airplane which had a custom made belly-hole for the camera lens on the floor. The digital images were stored in a PCMCIA card in DCS format (a Kodak proprietary format) during the flight and acquired into TIF files using Adobe Photoshop software. The acquired images had three 12 bits bands of G, R, and NIR with pixels and 4.4 MB size. The resolution of the camera varied with elevation of the aircraft. For example, at an elevation of 444 m the image resolution was 0.2 m/pixel and at 222 m elevation the resolution was 0.1 m/pixel, with a 20 mm lens. The images used in this study were taken at two elevations of approximately 180 m and 450 m. The 1997 images were taken using a 20 mm lens, but the 1998 images were taken using a 35 to 80 mm zoom lens. The FF farm images were geo-referenced to a spatial resolution of 0.6 m/pixel. The original resolution of the AGDE farm images used in this study was 0.2 m/pixel. CIR IMAGE ANALYSIS The images used for nitrogen stress studies were georeferenced to a resolution of 0.3 m/pixel; whereas, the images for yield prediction and soil analysis were georeferenced to a comparatively lower resolution. The combine s cutting width and hence the distance between adjacent rows of yield points in the yield map were 3 m for AGDE field, 4.5 m for FF field, and 6 m for WF field. A yield resolution of 3 m/pixel was chosen for yield modeling in AGDE and FF fields. A resolution of 6 m/pixel, which was same as the resolution of original yield map, was used in WF field for the ease of data handling since it was a large field (130 ha) compared to AGDE field (2.2 ha) and FF field (12 ha). For yield studies, both the yield and image files were converted to the same resolution of 3 m/pixel for FF and AGDE fields and 6 m/pixel resolution for WF field. The spatial variability in soil type was not as sudden as the variability in crop. Hence a lower resolution of 6 m/pixel was used for soil type studies. Histograms and three-dimensional image pixel distributions were used for the initial identification of the patterns in the image. The image pixel distribution was used to find out the number of clusters, that is, distinct and separable patterns, an image may have. All the images were clustered into several distinct groups using ISODATA procedure in MultiSpec software (Purdue University, West Lafayette, Ind.). The clustered images were digitized into ArcView shape files and overlaid with application maps, yield map and soil map for spatial and statistical analysis. The average spectral values (cluster centers) and variances within each cluster were the features of the clustered theme. SOIL TYPE ANALYSIS A CIR image of the AGDE soil was taken on 2 June 1998, when the field plant coverage was less than 5%. Clustered image of bare soil was compared to a detailed soil map of the AGDE farm developed using grid sampling for soil type analysis. Clustered images were overlaid with soil type map for comparing the similarity between the two. The soil map was converted into grids of 6 m/pixel resolution, and the soil type grid was geo-referenced and aggregated with the geo-referenced and clustered image of the field. The cluster patterns were classified into various soil types based on a maximum likelihood classification, in comparison with the soil map. The classified image was spatially analyzed with the soil map by comparing each pixel of the classified image with the soil type of the nearest soil grid point, using Euclidean distance. The accuracy of soil classification using clustered CIR images was assessed based on the percentage area that was correctly classified into a particular soil type. VOL. 42(6):

4 NITROGEN STRESS ANALYSIS For nitrogen stress analysis the CIR images were first clustered for better identification of the spatial patterns in the image. The image values were then averaged for each nitrogen plot in order to compare with the nitrogen rates. The clustered image was then geo-referenced and overlaid with maps of stalk nitrate, applied nitrogen, and spadmeter reading. The reflectance values were correlated to the applied nitrogen, and the stalk nitrate to find the potential of CIR images to identify nitrogen stress in corn. All the correlation analysis done in this article used Pearson s linear correlation method (Fruend and Wilson, 1993). Images were taken at approximately one-month intervals, for nitrogen analysis. YIELD MODELING For yield analysis, the yield map of each field was aggregated with clustered images and averaged over the clusters to obtain the average yield for each cluster. The algorithm for geo-referencing and theme aggregation was modified for averaging the yield within each cluster. The average yield per cluster was correlated to the average image bands values for each cluster as well as the spectral indices developed from the image bands. The spectral indices used in this study were normalized difference vegetative index [NDVI = (NIR R)/(NIR + R)], simple ratio index (SRI = NIR/R) and intensity normalized NIR (NNIR = NIR/I), intensity normalized R (NR = R/I) and intensity normalized G (NG = G/I), and normalized intensity (NI). Here, intensity (I) in the CIR range was taken as the sum of NIR, R and G, and NI was the intensity normalized by its maximum. Two sets of regression models were developed using stepwise selection procedure for each of the clustered images. The first set of models used the uncalibrated gray level values of each band of image as the independent variables. The second set used NDVI, SRI, NNIR, NR, NG, and NI. Two of the best models, one in uncalibrated spectral values and another in normalized indices were evaluated using images from other fields. The average prediction error per cluster, which is the percentage deviation of predicted yield from actual yield, was considered as the criterion for evaluating the model performance. RESULTS AND DISCUSSION SOIL TYPE PATTERNS The CIR image of AGDE soil showed some distinct patterns or spatial variability (fig. 1). The plant coverage over soil at this time was less than 5%, and the patterns were therefore assumed to be due to the variability in soil characteristics. Initial analysis of samples from soil image showed histograms with five to six distinct distributions. A sample histogram (fig. 2) showed five distributions in the NIR band and four well-defined distributions in both R and G bands of the spectral values. The spectral values from areas under the five major soil type in the field in figure 1 Figure 2 Histogram of a sample area from the color infrared image in figure 1. The near-infrared and green bands show five distinct distributions; whereas, the red band indicated four distributions. Figure 1 Color infrared image of a research farm, 17 days after planting. Lighter color indicates soil with comparatively less amount of clay TRANSACTIONS OF THE ASAE

5 Figure 3 Gray level distribution of color infrared image in figure 1 on NIR, Red, and Green coordinates. The five distributions correspond to the five soil types. were plotted with different legends, on coordinates of NIR, R, and G. This three-dimensional pixel distribution indicated five distinct and separable groups (fig. 3). These results provide strong evidence that the soil type was a major factor that contributed to the patterns observed in the soil image. The soil factors that had contributed to the different spectral values were mainly the clay content, Figure 4 A color infrared image of AGDE field after clustered (classification using an unsupervised learning method) into spatial patterns and overlaid with soil map. The overlay shows the similarities between a classified soil image and soil type map. Table 1. Result of soil type analysis using unsupervised classification of CIR image for AGEDE field* Classifi- Total Area Total Area Correctly cation Under Each Under Each Classified Accuracy Soil Type Soil (ha) Cluster (ha) Area (ha) (%) 148 Proctor silt loam Brenton silt loam Drummer silt clay loam Flanagan silt loam Throp silt loam Peotone silty clay loam Total * The overall classification accuracy is 76%. color of the soil and moisture content that is related to clay content under normal conditions. The soil map and clustered image showed strong similarity between the patterns (fig. 4). Using the overlaid themes, the area of each cluster that was correctly classified into a particular soil type was calculated along with the total area under each cluster (table 1). The CIR image was able to identify five of the six soil types. The accuracy of classification varied from 34% for Brenton silt loam to 98% for Thorp silt loam. Of the total area of the farm, 76% was classified correctly. However, the image processing technique did not identify peotone silty clay soil, which occupied a relatively small area of 0.4 ha in the field surrounded by drummer. Peotone could not be recognized from drummer soil because they both have the same amount of clay content and color (a value of 2 or 3, chroma of 1 or 2, and a hue of 10YR or neutral). The main difference between Peotone and Drummer soil is that they have different types of clay mixtures in them. The CIR imaging method did not identify some small patches of Brenton and Proctor silt loams. Proctor and Brenton are poorly drained soils like Drummer, but they are found in higher elevations compared to Drummer and are drier, differentiating them in the CIR image. At the time of imaging, the moisture content of the soil was very high throughout the field after extensive rain and flooding. This high moisture throughout the field made it difficult to identify some of the Proctor and Brenton soils from Drummer soil. Throp soil has lesser clay content as compared to Drummer soil and hence it appeared lighter in color. County soil map of the AGDE farm was very similar to the soil map used in this study. The results from CIR image analysis suggest that the image patterns identified from aerial remote sensing data under normal conditions (no flooding or crop) differentiated the soil types that are either at different elevations or have different surface layers that are distinguishable by color and clay content. This method may not be very accurate if the soil surface is covered with vegetation because the vegetation changes the spectral reflectance pattern. If the soil types have similar top layer but differences in their inner layers, the CIR imaging technique may not identify the differences. Classified CIR images of the soil would be a cheaper, easier and accurate guideline for soil mapping as compared to relying on visual analysis by experts. NITROGEN STRESS DETECTION Images acquired early in the season (first 8 weeks) for the nitrogen test plots in the FF field did not show any nitrogen stress. This was mainly due to the adequate supply of nitrogen in the early stages of crop growth. The nonitrogen plots started showing up the stress patterns in the images from 75 DAP onwards. As indicated by the later images, stress developed in the 67 kg/ha (60 lb/a) nitrogen plots at a much later time. After clustering, the CIR image of the nitrogen plot in FF field taken on 21 August 1998 (125 DAP) showed the low nitrogen plots distinctly separate from the rest of the plots (fig. 5). The 3-dimensional distribution of reflectance from different nitrogen plots in the FF field on 125 DAP showed nitrogen stressed and well nourished plants as separate groups (fig. 6). The three groups of no-nitrogen, 67 kg/ha nitrogen and well nourished (greater than 134 kg/ha nitrogen) VOL. 42(6):

6 Figure 5 A color infrared image of the nitrogen plot taken on 21 August 1997, after classifying into groups using ISODATA clustering. Table 2. Results of nitrogen analysis for FF and AGDE field* Average Gray Level Value Applied Stalk Spad of CIR Image Bands Nitrogen Nitrate Reading Field Plot (kg/ha) (ppm) of 24/7/98 NIR R G FF A FF B FF C FF D FF E FF F FF G FF H FF I FF J AGDE AGDE AGDE AGDE AGDE AGDE AGDE AGDE AGDE * The gray level values for FF field are from 21/8/97 image and that for AGDE farm are from 31/7/98 image. --- Stalk nitrate was below detection limit. - Spad reading are not available for FF field. Figure 6 Gray level distribution of color infrared image of nitrogen plot in FF field on NIR, Red, and Green coordinates. Different colors indicate nitrogen levels. formed three overlapping groups. The overlapped groups indicate that it is possible to separate nitrogen stress areas but not with 100% accuracy. The high nitrogen levels could not be differentiated among themselves. The three-band image values showed a negative trend with respect to applied nitrogen, and stalk nitrate, for all the images of FF field (table 2). The low correlation of different nitrogen levels with NIR and spectral indices derived using NIR, for images from 2 July to 21 August of 1997 was due to the saturated NIR band (table 3). Due to the heavy rainfall and flood in 1998 crop season, the trends in stalk nitrate and spad reading showed little resemblance to the applied nitrogen. In some plots in AGDE field, stalk nitrogen was less than the detection limit of 25 ppm. As expected, both R and G bands were negatively correlated to the applied nitrogen level since more and more of visible spectrum was absorbed by the increasing amount of chlorophyll pigment. The earlier images of FF field did not show any nitrogen stress since the nitrogen was sufficient in the beginning, to meet the requirement of the germinating plants. However, nitrogen stress started showing up approximately two months after planting. It was difficult to separate plant signature from soil background at low canopy coverage of less than 10% in the earlier stages of plant growth. At such low canopy coverage, the noise due to soil background dominated the spectral values sensed by the remote sensing system. A maximum correlation coefficient of 0.78 was observed on 125 DAP between applied nitrogen, and R and G values, but it dropped down as the crop matured. This drop in correlation was due to the substantial wind damage to the crop before the last two sets of images were taken. Any difference in the metabolic activity, crop water content, leaf area index, crop orientation, shadowing or chlorophyll content will change the crop reflectance pattern (Bauer, 1985; Knipling, 1970). The partially lodged crop of the FF field had a disrupted crop orientation, leaf area index, and stress that would have contributed to a difference in the spatial reflectance pattern of the canopy. This could be the reason for a reduced correlation between nitrogen levels and spectral values after the wind damage. NDVI and SRI were better or equally correlated to applied nitrogen than the gray level values of the image bands. These results agree with the findings of Ma et al. (1996). The correlation between stalk nitrate and spectral values was not significant during the active growing period (table 3). The stalk nitrate was tested at the time of harvest. The non-significant correlation was due to the depletion in stalk nitrate during senescence and the effect of other factor like wind damage and change in plant tissue color during senescence. However, spadmeter readings were highly correlated to R and G bands of the image, and the spectral indices. YIELD MODELING The spatial variation in yield (fig. 7) was similar to the spatial variation found in CIR images of the field (fig. 5). The variation of yield with image bands (fig. 8) showed 1916 TRANSACTIONS OF THE ASAE

7 Table 3. Correlation coefficient between nitrogen levels (applied nitrogen, stalk nitrate and spad reading) and CIR spectral values Field Date NIR R G NNIR NR NG NDVI SRI NI Correlation Coef. Between Applied Nitrogen Level and CIR Image FF 13/06/ FF* 02/07/ FF* 26/07/ FF* 21/08/ FF 06/09/ AGDE 02/07/ AGDE 31/07/ AGDE 21/08/ Correlation Coef. Between Stalk Nitrate and CIR Image FF 13/06/ FF* 02/07/ FF* 26/07/ FF* 21/08/ FF 06/09/ Correlation Coef. Between Spadmeter Reading and CIR Image AGDE 31/07/ * NIR was saturated. The correlation coefficients in italics are not significant. negative linear trends for NIR, R, and G bands for AGDE field. It means that the high yielding clusters had low R and G values. The G band showed minimum relative decrease with increase in yield. However, the variation of NIR with yield was positive for FF field images. Both R and G bands were better correlated to yield with a correlation coefficient in the range of 0.84 to 0.99 as compared to the NIR band (fig. 8 and table 4). The correlation of image bands with yield was strongly linear and this agrees with the results of Senay et al. (1998). The negative trend of R and G with yield is understandable since the visible light reflectance decreases with better plant health due to increased pigment content in healthy plants as compared to stressed plants. The lack of correlation and negative trend between NIR and yield for AGDE field may be due to the herbicide plots in half of the field area. The yield from herbicide plots was considerably less though there was a good amount of biomass in the field due to the presence of weeds, which enhanced the NIR reflectance. The inconsistent trend between NIR and yield in FF field is due to the saturated NIR band in some of the 1997 images. The correlation was high and significant after pollination, for images from 21st August onwards, compared to the images before pollination. The NDVI, SRI, and intensity normalized image bands were strongly correlated to yield for images after pollination. NI was the most correlated index among all the indices considered in this study. NI was negatively correlated to yield since visible light constituted 66% of NI and visible light reflectance was negatively related to plant health. In WF field two distinct and parallel trends were observed between NI and normalized yield (fig. 9). The ground truth revealed that these two trends represented the areas that were regular planted (WF1) and replanted (WF2) after the rain damage. These two trends were correlated separately to the yield. The replanted areas showed a very strong and linear correlation compared to the rest of the field. The high linear correlation between various image indices and yield showed that the spatial variation in yield was indicated by the spatial patterns observed in the CIR image and yield can be modeled using the CIR images of the field after pollination. The initial yield modeling was done using data from FF field. A CIR image acquired on 125 DAP (21 August 1997) Figure 7 Yield map of FF field from 1997 overlaid with application map, and a clustered color infrared image converted into polygons. The legend shows yield in bushels/acre. (a) (b) Figure 8 Relationship between the gray level values of NIR, red (R), and green (G) band of CIR image and yield for (a) FF field on 6 September 1997, and (b) AGDE field on 21 August VOL. 42(6):

8 Table 4. Correlation between yield and spectral indices derived from CIR images, for different experiment fields (for WF field, WF1 corresponds to regularplanted areas and WF2 corresponds to replanted areas) Field Date NIR R G NNIR NR NG NDVI SRI NI Correlation of Yield with Different Image Indices FF 13/06/ FF* 26/07/ FF* 21/08/ FF 06/09/ AGDE 02/07/ AGDE 31/07/ AGDE 21/08/ WF1 21/08/ WF2 21/08/ * NIR component was saturated. The correlation coefficients in italics are not significant. was used to fit a spatial yield model from the image bands since this image showed maximum correlation with yield among all the images for FF field. Stepwise regression using uncalibrated image bands of NIR, R, and G resulted in a linear model on R (yield = R) with an R 2 value of A second-degree polynomial model on R (yield = 0.07 R R 1762) resulted in an R 2 value of No other spectral band or combination of spectral band met the significance criterion (p-value < 0.15) to find entry into the model. The yield predicted using the polynomial model plotted against the actual yield showed a good prediction, with a slope of 0.96 and intercept of 6.36 (fig. 10). It should be noted that 1997 was a normal year and the wind damage towards the end of the season had little effect on the yield. Testing the polynomial and linear models developed for FF field from image on 21 August 1997 in other fields and images showed that the model does not work for a different field or season. Stepwise regression of yield was performed on each field and image separately, using uncalibrated image bands. The resulting models were very different for each field (table 5). For example, the model for AGDE field had all three spectral band values in the model while the model for WF field had only G value in the model; whereas, the FF field had R in the models before wind damage and G in the models after wind damage. All the models showed very good R 2 values in the range of 0.76 to These good fits of the spatial models indicated that the spatial patterns within a CIR image represented 76 to 98% of the spatial variation in spatial yield. The variation in light intensity, the difference in relative reflectance of different hybrid varieties and ground, and atmospheric conditions were the reasons why yield models on uncalibrated image bands were unique for each field. Hence, a second stepwise regression analysis was performed on normalized spectral indices derived from uncalibrated spectral values. Stepwise regression analysis on normalized image indices resulted in linear models on NI (table 5) with R 2 values varying from 0.6 for FF field to 0.99 for WF field. None of the other spectral indices met the significance level (15%) for entry into the model. In WF field, the two distinct trends corresponding to regular and replanted areas were analyzed separately. The regression lines for these two areas were parallel with different intercepts. This result indicated that for different planting dates, the variation in yield with respect to normalized intensity was similar but the replanted crops reflected relatively high intensities of light than the older crops. The yield model on uncalibrated image band values developed from 21 August 1997 image of FF field and the model on normalized intensity developed from 21 August 1998 image of AGDE field were tested in all the three fields. As expected, the model on uncalibrated gray level values resulted in high prediction error of over 100% (table 6). The band values of image varied quite significantly depending on sunlight intensity and hence, a model on uncalibrated gray level values was not reusable. The model on normalized intensity also had a high error of Figure 10 Yield predicted from uncalibrated image band values from 21 August 1997 plotted against actual yield. Slope = 0.96, intercept = Figure 9 Variation of normalized yield with respect to normalized intensity for WF field in Normalized intensity was calculated for a CIR image from 21/8/98. Trend 1 represents all the regularplanted areas and trend 2 represents replanted areas after the rain damage in the beginning of the season. Table 5. Results of regression analysis using stepwise selection procedure, between yield and spectral indices for different fields* Yield Model R 2 Yield Model R 2 Field Date on Intensity Value on NIR, R, and G Value FF 02/07/97 ny = NI 0.79 Y = R 0.76 FF 21/08/97 ny = NI 0.85 Y = R 0.89 FF 06/09/97 ny = NI 0.60 Y = G 0.77 AGDE 21/08/98 ny = NI 0.87 Y = NIR R 21.2 G 0.91 WF1 21/08/98 ny = NI 0.85 Y = NIR 0.42 R 0.92 WF2 21/08/98 ny = NI 0.99 Y = R 0.98 ny = normalized yield, NI = normalized intensity, Y = yield, NIR = near infrared, R = red, and G = green. * In WF field the regular (WF1) and replanted (WF2) areas were modeled separately TRANSACTIONS OF THE ASAE

9 Table 6. Results of the evaluation of two yield models on different fields and images* Average Prediction Error per Cluster of Model in Different Fields and Images (%) FF FF FF WF1 WF2 AGDE Model Tested 26/07/97 21/08/97 06/09/97 21/08/98 21/08/98 21/08/98 Y = R ny = ni * Model for yield (Y) was developed using FF field data from 21/8/97 and the model for normalized yield (ny) was developed using AGDE field data from 21/8/ to 45% except for the replanted areas in WF field. The normalized model was developed using AGDE image data and, hence, the prediction error for this field was the lowest at 9%. The results showed the potential for developing a yield prediction model from CIR images of the field. CONCLUSIONS The histograms and three-dimensional pixel distribution of the CIR images of a crop field showed clear grouping of pixels, indicating the spatial variability within the field. These spatial patterns resembled to the predominant factor (soil or crop) at the time the image was acquired. A spatial analysis of the soil image showed that the image patterns strongly resembles the soil type. The soil image after an unsupervised classification identified six of the seven soil types and accurately classified the soil type with an average accuracy of 76% as compared to a conventional soil map. Aerial CIR images of bare soil could replace the more expensive and laborious traditional soil mapping technique used by trained professionals. CIR images also can be used as an accurate guideline for soil sampling in the field instead of relying on visual analysis since CIR images can pick up patterns that are not very visible to human eye. The CIR aerial images could delineate the nitrogen stress areas in the field. However, differentiating between the nitrogen levels of well-nourished areas was difficult. The CIR reflectance values were significantly correlated to the applied nitrogen and spadmeter reading. The normalized difference vegetative index and simple ratio index were better indicators of nitrogen stress than the uncalibrated image gray level values. The CIR reflectance was well correlated to the yield after pollination of corn compared to the early images. The relationship between the image band values and yield was strongly linear. R and G bands showed better correlation with yield as compared to NIR band. Yield showed negative correlation with R and G, and positive correlation with NIR for fields under total herbicide application. However, yield was negatively correlated to NIR in fields infested by weed. Modeling of yield on uncalibrated image bands showed that 76 to 98% of spatial variation in yield could be explained by the spatial patterns visible in CIR images. Because of variable lighting conditions and different crop varieties, the yield prediction models on uncalibrated image values were not exchangeable among fields or seasons. A linear yield model on NI predicted yield in different fields with a prediction error below 45%, which was better than the model on uncalibrated image values. Further research is required to modify the model for better performance and higher resolution. For example, incorporation of additional factors like moisture level, and variety will improve the model performance. This study used average reflectance values and yield for clusters obtained from unsupervised classification of CIR images. A high-resolution yield analysis at pixel levels instead of analyzing relatively fewer clusters obtained from image classification will be more accurate and meaningful. ACKNOWLEDGMENT. The authors express their gratitude to Mr. John Beal for his assistance in field data collection. The authors also thank the Illinois Council of Food and Agricultural Research for supporting this project (No AE). The use of trade names, proprietary products, or specific equipment is only meant to provide specific information to the reader, and does not constitute a guarantee or warranty by the University of Illinois, and does not imply the approval of the named product to the exclusion of other products that may be suitable. REFERENCES Adcock, T. E., F. W. Nutter, and P. A. Banks Measuring herbicide injury to soybeans (Glycine max) using a radiometer. Weed Sci. 38(6): Bauer, M. E Spectral inputs to crop identification and condition assessment. Proc. IEEE 73(6): Bausch, W. C., and H. R. Duke Remote sensing of plant nitrogen status in corn. Transactions of the ASAE 39(5): Blackmer, T. M., J. S. Schepers, and G. E. Meyer Remote sensing to detect nitrogen deficiency in corn. In Site-Specific Management for Agricultural Systems, eds. P. C. Roberts, R.W. Rust, and W. E. Larson, Madison Wis.: ASA/CSSA/SSSA. Chappelle, E. W., M. S. Kim, and J. E. McMurrtrey III 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): Curran, P. J Aerial photography for assessment of crop condition: A review. Appl. Geogr. 5(4): Filella, I., L. Serrano, J. Serra, and J. Penuelas Evaluating wheat nitrogen status with canopy reflectance indices and discriminant analysis. Crop Sci. 35(5): Fruend, R. J., and W. J. Wilson Statistical Methods. SanDiego, Calif.: Academic Press. Gopalapillai, S., L. Tian, and J. Beal Detection of nitrogen stress in corn using digital aerial imaging. ASAE Paper No St. Joseph, Mich.: ASAE. Heute, A. R A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25(3): Jackson, R. D., P. J. Pinter, R. J. Reginato, and S. B. Idso Detection and evaluation of plant stresses for crop management decisions. IEEE Trans. Geosci. & Remote Sens. GE-24(1): Knipling, E. B Physical and physiological basis for the reflectance of visible and near-infrared radiation from vegetation. Remote Sens. Environ. 1(3): Ma, B. L., M. J. Morrison, and L. M. Dwyer Canopy light reflectance and field greenness to assess nitrogen fertilization and yield of maize. Agron. J. 88(6): Moran, M. S., Y. Inoue, and E. M. Barnes Opportunities and limitations for image based remote sensing in precision crop management. Remote Sens. Environ. 61(3): Penuelas, J., J. A. Gamon, A. L. Fredeen, J. Merino, and C. B. Field Reflectance indices associated with physiological VOL. 42(6):

10 changes in nitrogen- and water-limited sunflower leaves. Remote Sens. Environ. 48(2): Qi, J., A. Chehbouni, A. R. Huete, Y. H. Kerr, and S. Sorooshian A modified soil adjusted vegetation index. Remote Sens. Environ. 48(2): Sawyer, J. E Concepts of variable rate technology with considerations for fertilizer application. J. Prod. Agric. 7(2): Senay, G. B., A. D. Ward, J. G. Lyon, N. R. Fousey, and S. E. Nokes Manipulation of high spatial resolution aircraft remote sensing data for use in site-specific farming. Transactions of the ASAE 41(2): Tian, L., R. Hornbaker, and R. Schmidt Aerial field sensing and mapping for precision farming. Paper No. AA Presented in NAAA/ASAE Joint Technical Session, Las Vegas, Nev. St. Joseph, Mich.: ASAE. Tomer, M. D., J. L. Anderson, and J. A. Lamb Assessing corn yield and nitrogen uptake variability with digitized aerial infrared photographs. Photogram. Eng. & Remote Sens. 63(3): TRANSACTIONS OF THE ASAE