Use of Remote Sensing to Determine Plant Health

Use of Remote Sensing to Determine Plant Health and Productivity Chuan-Shin Chong1, John P. Basart1, Forrest W. Nutter2, Jr., Gregory L. Tylka2, Jie Guan2 1

Department Of Electrical Engineering, Iowa State University 2 Department Of Plant Pathology, Iowa State University

ABSTRACT This project seeks to assess plant productivity and health in time and space by measuring spectral reflectance from soybean canopies using remote sensing images that do not require ground assessment. Aerial images and reflectance measurements from a multi-spectral radiometer were obtained simultaneously from a soybean field located in Story County, Iowa. The multi-spectral radiometer has eight wavelength bands, ranging from 460-nm to 810-nm and was used as a ground reference for the data analysis. Aerial images were obtained from altitudes ranging from 152 to 427 meters from the ground during summer 2000. Aerial images were analyzed using Matlab, ArcView and Imagine. Difficulties in image analysis and interpretation may occur as the sensing equipment increases in altitude because atmospheric influences become more pronounced. Scattering and absorption of electromagnetic waves in the atmosphere change the spectrum of the reflected wave emitting from the plants as it propagates from the plants to the sensors. Color calibration procedures were used with red, green and blue ground cloths to correct aerial images in the respective red, blue and green bands. Regression analysis was carried out to quantify the relationships between multi-spectral radiometer data and aerial image data. Keywords: soybeans, yield, productivity, remote sensing, aerial images, radiometer, infrared, regression, calibration and classification.

1. INTRODUCTION Yield and plant health can be related to the quality and amount of sunlight reflected from crop canopies (Nutter et al, 1999 and Nutter et al, 2000). Chlorophyll and auxiliary pigments in plants are sensitive to radiation. Every crop canopy has its own spectral characteristics, which reflect different amounts of radiation as a function of wavelength. In general, higher reflectance for crop canopy occurs in the near-infrared band (Lillesand, 1999). Thus, the purpose of this project was to find a relationship between soybean health and yield using images of specific reflectance wavelengths to improve efficiency of crop health assessments. In addition to aerial images, remote sensing experiments using a hand-held multispectral radiometer were also carried out in order to test for the correlation at different levels of scale (Nutter et al, 2000). The radiometer was nonimaging and combined all data within a 40o field of view. The radiometer technique is limited by height of about 2.5 m and is not feasible if a field is large. A method with higher altitude is necessary to improve efficiency. To this purpose we conducted a number of imaging observations with an aircraft flying between heights of 152 to 427 meters. The current maximum height of 427 meters allowed the images to cover the whole the experimental soybean field (119 meters x 102 meters), with a 30-degree view angle lens. The advantage of images compared to a multi-spectral radiometer is that simultaneous observation of a large area can be obtained. Imaging equipment can also be attached to advanced hardware or technology to enhance the functions. Due to the requirement of visiting the field, radiometer readings may not be obtained if accessibility to the field is bad. By using aerial images, plant condition data can be obtained from large areas that are difficult for human access regardless of ground conditions. High speed post-processing by a personal computer is an advantage for image processing. An image can contain 1000 times more data than that obtained by the radiometer at one time. Only a small area can be covered by the radiometer at each assessment, whereas images taken from an aircraft can cover large crop areas.

484

Infrared Spaceborne Remote Sensing IX, Marija Strojnik, Bjørn F. Andresen, Editors, Proceedings of SPIE Vol. 4486 (2002) © 2002 SPIE · 0277-786X/02/$15.00

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/07/2015 Terms of Use: http://spiedl.org/terms

2. EXPERIMENTS AND RESULTS 2.1 Equipment The equipment used in the experiment was a hand-held multi-spectral radiometer (Cropscan Inc., Rochester, MN), Two 35 mm single lens reflex ZX- 50 Pentax cameras, one with a 810 nm wavelength filter, SMCP-FA 50 mm f/1.7 lens with view angle 47º, Kodak Gold 200 daylight film, Kodak high speed infrared film (HIE 135-36) and 1978 C-152II Skyhawk aircraft. The multi-spectral radiometer has 8 filters: 460 nm with bandwidth of 6.8 nm, 510 nm with bandwidth of 7.7 nm, 560 nm with bandwidth of 9.4 nm, 610 nm with bandwidth of 10 nm, 660 nm with bandwidth of 12 nm, 710 nm with bandwidth of 12 nm, 760 nm with bandwidth of 11 nm and 810 nm with bandwidth of 11 nm. The images were only taken in the visible band and the 810 nm infrared band with bandwidth of 11.4 nm. 2.2 Data Collection Images were taken at six different altitudes at 152 m, 183 m, 244 m, 305 m, 366 m and 427 m during each flight. The cameras were inserted through a hole in the bottom of the aircraft. Data collection was planned every two weeks from plant date to harvest date. However, due to weather conditions and aircraft availability, the schedule occasionally was changed. The sun angle is also a concern in aerial photography. When the angle is not normal to the ground, it can change the information of the image. The effect will be more obvious when the sun is on the horizon. In order to avoid the sun angle problem, aerial photography was carried out between 11:00 am to 3:00 pm. Table 1 shows the schedule of the data collecting flights: Obsevation No. 1 2 3 4 5 6

Date

Weather Condition

May 19 May 22 June 15 June 29 July 7 July 13

Mostly Sunny Mostly Sunny Partly Cloudy Partly Cloudy Drizzle / Mostly Cloudy Mostly Sunny

Observation No. 7 8 9 10 11

Date

Weather Condition

August 2 August 9 August 25 September 8 September 18

Mostly Sunny Partly Cloudy Partly Sunny Partly Sunny Mostly Sunny

Table 1 Observation dates and weather conditions.

The soybean field was 119 meters x 102 meters in dimension. In order to study the field, it was divided into a "checkerboard" map with 30 quadrats x 49 quadrats. Soybeans were only planted in the quadrats which appear as light areas in a checkerboard pattern (see Figure 1). Each quadrat had four rows of soybeans. Four radiometer readings were obtained for each quadrat by manually walking along the soybean rows with the radiometer held two meters above the ground. The other advantage of dividing the field into quadrats is that we can efficiently compare radiometer readings with image information. 2.3 Digitalization After acquiring the aerial images, the films were sent to a commercial photo center to develop into 4 by 6 inch prints. Then the prints were scanned using a Hewlett Packard 5100C Scanner into digital format. All the files were saved into JPEG format with zero compression. 2.4 Infrared Images The planting date for summer 2000 was May 23, 2000. From Figure 1, very little vegetation can be seen on June 15, 2000 (dark image) even though it was 23 days after the planting date. The intensity on June 15 was low because bare soil still dominates most of the field. Soil has lower reflectance in the infrared band compared to vegetation. Some higher intensity can be seen on the image because of corn debris from the previous year. The black spot that appeared in the upper right on August 25 was a sudden incident as it happened in a very short period. The plants were burned and dry. The hypothesis for the cause of the incident is lighting as virus or bacteria would need a longer period to cause damage to the plants. However, no strong evidence is available to prove the hypothesis. The condition of the plant can be measured by the amount of green leaf area because disease injury caused to crops reduces the healthy green leaf area (Nilsson, 1995 and Nutter, 1990). The higher amount of green leaf area reflects higher radiation in the near infrared

Proc. SPIE Vol. 4486


485

spectrum (Sanchez, 1999) compared to the yellow and dry leaves because of the green pigment, chlorophyll, contained in the leaves. In Figure 1, it shows that a higher intensity or reflectance location is the location that has more green leaf area (healthier plant). The small black square box between two small white boxes in the figure are the blank soil buffer which contains no soybeans.

June 15, 2000

August 2, 2000

August 9, 2000

August 25, 2000

September 8, 2000

September 18, 2000

Figure 1. Woodruff field (Story County, Iowa) infrared images in summer 2000. The darker shading in the images indicates less healthy plants. 2.5 Calibration Calibration for high altitude images is needed because there are distortions in the atmosphere such as dust, small particles, smoke, water vapor and etc (Hadjimitsis et al, 2000). These distortions can cause a change in the intensity of each band. The materials used to calibrate the high altitude images were red, green and blue cloths. Ground images of the cloths served as reference wavelengths for the aerial images. This method calibrates intensities of three different band layers. A combination of different intensities of the three bands produces different colors. In this method, the distribution of the distortion was assumed constant throughout the area captured by the camera. Calibration procedure is shown as follows:

Extract cloth images

Find red, blue, green factors

Calibrate Image

Input image

Figure 2. Diagram shows the steps to calibrate the images.

Using the ground reference cloth images and aerial images, a correction factors for each of the three bands was obtained as follows:

486



Factor (Red Band), FR =

I R ,G

Factor (Green Band), FG = Factor (Blue Band), FB =

(1)

I R, A I G ,G

(2)

I G, A I B ,G

(3)

I B,A IR,G = Red band average intensity of ground red cloth image (scalar) IR,A = Red band average intensity of aerial red cloth image (scalar) IG,G = Red band average intensity of ground green cloth image (scalar) IG,A = Red band average intensity of aerial green cloth image (scalar) IB,G = Red band average intensity of ground blue cloth image (scalar) IB,A = Red band average intensity of aerial blue cloth image (scalar).

Using the factors and the images date we create the matrices. F = [FR FG FB ] and I = [I R I G I B ] where IR = Two-dimensional matrix of red band intensity IG = Two-dimensional matrix of green band intensity IB = Two-dimensional matrix of blue band intensity. The correction to the image is accomplished by the operation Y = FI

(4)

where Y = calibrated image. Before doing the region selection on the image, geometric distortion of the image was corrected. The distortion was caused by the line normal to the camera not being perpendicular to the ground due to the pitch and roll of the aircraft. The correction was done using Paint Shop Pro and ArcView software. The tools used were the image deformation in Paint Shop Pro and alignment function in ArcView which can select ground control points and rectify the image. 2.6 Region Selection Every quadrate on the image was selected using ArcView. A grid map was created to match the quadrats in the field. The average intensity for each band for each quadrate was then calculated. So that, every box of the grid map could be represented by one value. The values were used later for data processing and data comparison processed. The blank soil buffer on the image was not taken as part of the data. The total quadrats obtained were 995. 2.7 Radiometer and Image Data Comparison The next step in the processing was to compare the images to the radiometer data. From the visible images, broad band red ,green and blue images were extracted. Due to the limited number of bands for the images, only four bands were used for the comparison. The 660 nm radiometer band was selected to compare with the red image band, 560 nm was selected to compare with the green image band and 460 nm was selected to compare with the blue image band. After the regions on each image were selected, the average value of the 810 nm intensity from each quadrat was calculated and compared to the average value of the radiometer data over the quadrat. The data were then plotted as shown in Figure 3. There are total of 995 points in each plot. The figures are the sample patterns of the distribution of the points on two dates, August 25 and September 8, 2000. The R2 value obtained on August 25 data after applying linear regression is 0.8267. R2 for the September 8 data is even better. The value is 0.9121. This value can interpret that 91% of the radiometer data can be explained by the image data. The R2 values for the rest of the flights are shown in Figure 4.



487

Image Intensity

Percentage Reflectance (Radiometer)

Percentage Reflectance (Radiometer)

Linear Regression between Radiometer Data and Aerial Image data, Aug 25, 2000

Linear Regression between Radiometer Data and Aerial Image data, Sep 8, 2000

Image Intensity

Figure 3. Relationship between radiometer data and image data There is a big gap between day 167 and 215 in Figure 6 because the IR films were accidentally destroyed by the photo service shop while in the process of developing. However, based on the data collected in summer 2000, the coefficients of determination are very high and we can conclude that the relationship is very strong between the radiometer and aerial images. The relationship is expected to improve further by calibrating the infrared images. Calibration for infrared film will be carried out in future work.

Image Data lost

Figure 4. R2 value for aerial IR images and radiometer data. 2.8 Yield The yield data were also used in the comparison to find the relation with the bands. For the yield data, soybeans were harvested according to quadrats. The soybeans for each quadrat were dried. The temperature of the dryer was set to 80° F and the plants were dried for 3 days until the moisture value dropped to 5 or 6 %. The beans were then weighed. Before the yield classification was carried out, the image data was scaled with the season coefficient, k calculated as k=

488

max(radiometer data) . 255



(5)

k was calculated by using the highest radiometer intensity at 810 nm for the entire field. Six values of k obtained from the summer 2000 data (one point for each observation day) were plotted in Figure 5. A second polynomial was fit to the six points. Smooth k values for the whole season can then be obtained from the polynomial line fit. The season coefficients were low at the beginning and the end of the season as the percentage of the reflectance was low according to radiometer data.

August 25, 2000

k = −4 x10 −5 t 2 + 0.00625t − 0.0772

Days After Planting Date, t

Data Count Figure 5. Second order polynomial fit for season coefficient,

Figure 6. Decision boundaries (horizontal lines) for three

Woodruff Farm Woodruff Farm 2000

Woodruff Woodruff Farm 2000

Woodruff Farm 2000

10

10

0.8

80

0.6

60

60 Co effi cie 40 nt

Coefficient of Determination

80



1

Co effi cie40 nt of det20 er mi

0.4 0.2 0

of det er 20 mi nat 0

0

-0

120 140 160 180 200 220 240 260 280

-20 12 140 14 160 16 120

Day of

Day of Year

DayDay of Year of Yyear

18 200 20 220 22240242602628028 180 Day of

Dayofofyear Year Day

Figure 7. The coefficients of determination were obtained from the linear regression of yield and radiometer data (left), yield and image data (middle) and yield and NDVI (right).



489

The entire yield was divided into three classes. Class 1 (low yield) contains the yield in a quadrat from 1 to 380 g, class 2 (medium yield) contains the yield from 381 to 768 g while class 3 (high yield) contains yield from 769 to 1140 g. The classification using fuzzy method (Wang, F., 1989, 1990a, b, 1991) was also applied to the yield (Chong, 2001). The advantage of using a fuzzy method is that it includes the sub-pixel analysis. The result of the classification can be expressed in the membership scale from 0 to 1. Four bands (red, green, blue and IR bands) were used in the fuzzy classification. Fuzzy mean and covariance matrices are the main components to carry out the analysis in the fuzzy classification. How well the yield, radiometer, image and NDVI data can fit in the linear regression was assessed (See Figure 7). The pattern of the plots is very similar. Yield has a very strong linear relationship with radiometer, image and NDVI data at the middle of the season. The relationship is weak at the beginning of the season as the plants were hardly seen from high altitudes. Figure 8 shows the result of the classification using a fuzzy system. As mentioned earlier, the advantage of using fuzzy system is the sub-pixel analysis.

Membership Grades of Yield Class 1



Figure 8. Membership grades for 3 classes on August 25 data

The values of the membership grades are expressed in images with a grayscale from 0 to 255 scaled to 0 to 1 because the membership grades range from 0 to 1. Areas with value 1 or pure white means the area has the highest characteristic of the class whereas value 0 or pure black represents the lowest characteristic. From the three classes, we can see the gray color of some quadrats, which is in middle range (0.3-0.7) of the membership scale. However most of the quadrats' values are either extremely towards value 0 or extremely towards value 1. Gray color in middle range can only be seen at the boundary between classes. This is because the training data was obtained in pure membership grades (manually selected from raw data with a membership grade equal to one) but in reality they were in mixture of several classes. From the results, we can see that three classes were separated very relatively well. For low yield class (class 1), not many areas have really low yield. However, the medium class (class 2) has the most areas compared to other classes. High yield areas for class 3 only occur at certain parts of the field such as the lower right and left corners. However, the result was not really satisfactory due to the broad visible bands on the images.

Data Actual Yield Data IR Images (August 2, 2000) IR Images (August 9, 2000) IR Images (August 25, 2000) IR Images (September 8, 2000)

Number of Quadarts in each Classification 1 - 380 g 381 - 768 g 769 - 1140 g (Low Yield) (Medium Yield) (High Yield) 64 752 179 2 322 671 24 384 587 25 390 580 51 778 166

Table 2 Boundary classification results

490



The boundary classification was then carried out for the IR band. Decision boundaries were made for all 3 classes on August 25 data. The August 25 data was picked as the determination coefficient between yield and radiometer data was the highest. The boundaries were used for August 2, August 9 and September 8 data. A plot was made (Figure 6) of the three classes. The abscissa is a running tall of each point in a class. The numbering for each class started at "data count" 1. A boundary is the average value of the highest value from lower class and the lowest value from the higher class (Figure 6). The values lying below the lower boundary were classified in class 1, the values lying above the lower boundary and below the upper boundary were classified in class 2 and the values lying above the upper boundary were classified in class 3. Table 2 shows the result of the boundary classification of 4 dates. The number of the quadrats in the classes changed during the season. From the result (Table 2), we can see a trend in the low, medium and high yield classes. More and more quadrats fell into the low yield class and medium class when the date was approaching the harvest date. Many of the quadrats switched from high yield to medium yield class on September 8. This might be due to the growth of the plants. The predicted high yield areas had a fast growth at the beginning of the season and caused the quadrates to fall in the high yield category at the beginning of the season. However, the growth reduced at the end of the season and the yield obtained was not as expected. Thus, many quadrats fell into the medium yield class. The predicted locations for low, medium and high yield were quite accurate. This can be seen from Figure 9. The health of the plant is highly related to the yield. A healthy plant will produce relatively high yield at the end of the season according to the structure and weather condition of the location. If the low yield can be known, scouting or field observations can be carried out to assess the plant health condition.

4. CONCLUSION The 810 nm IR band from images have high correlation with 810 nm radiometer data. The correlation varies from time to time in a season. The highest correlation happens during the plants mature season. With the information from aerial images, hopefully the radiometer data can be replaced with a higher efficiency method. The classification method would be helpful to farmers in determining which part of their fields might cause the low yield and immediate action or scouting can be taken to improve the health of the plant and increase the yield.

5. FUTURE WORK An automatic imaging system is in the progress of development. The system is designed to provide guidance and control for hardware and software to capture specific photographs of land features during flights. An aircraft will carry the device. The system will have the capability of receiving GPS coordinates, overlaying the GPS coordinates on recorded video images of a flight, relaying the GPS coordinates serially to a laptop computer, and controlling a set of still-frame cameras used to capture visible and infrared spectrum photographs. Both GPS and DGPS data will be gathered to increase resolution. This will allow for a more precise designation of points that are above the desired region to be photographed. The equipment involved in the system are GPS-DGPS receivers, still cameras, video camera, video digital overlay (VDO), monitor, VCR and camera mounts.

6. REFERENCE 1.

2.

3. 4.

Chong, Chuan-Shin, J. P. Basart, F. W. Nutter, G. L. Tylka, J. Guan,and C. C. Marett, “Determining the Health and Productivity of Crops Using High Altitude High Altitude Images,” to appear in American Society for Photogrammetry & Remote Sensing (ASPRS) 2001 proceeding, 11 pages, 2000. Hadjimitsis, D.G., Clayton, C.R.I, Hope, V.S., "The Importance of Accounting Atmospheric Effects in Satellite Remote Sensing: A Case Study from the Lower Thames Valley Area, UK," Space 2000, American Society of Civil Engineers, pp. 194-201, 2000. Lillesand, Thomas M., Kiefer, Ralph W., Remote Sensing and Image Interpretation. John Wiley & Sons, Inc., pp. 15-17, 1999. Nutter, F. W. Jr., "Remote Sensing and Image Analysis For Crop Loss Assessment," Crop Loss Assessment in Rice, International Rice Research Institute, Los Banos, Philippines, pp. 93-105, 1990.



491

Yield Class 1

Yield Class 2

Yield Class 3

Figure 9. Yield Classification Map using the boundary method. Each white square is a classified quadrat in its actual location in the field. The first row is for August 2, 2000, the second row is for August 9 (second row), the third row is for August 25, the forth row is for September 9 and the fifth row is for actual yield map.

492



5. 6.

7.

8.

9. 10. 11. 12. 13.

Nilsson, H. E., "Remote sensing and image analysis in plant pathology," Annual Review of Phytopathology. 15:489527, 1995. Nutter, W. Forrest, Guan, J., Percifield, R., Menelas, Blucher, Rosburg, T, Bidney, Sara, Basart, John P., Chong, C.S., Mike, C., "Use of Remote Sensing to Quantify the Impact of Plant Disease Epidemics on Plant Biomass and Productivity," in Ninth Annual Iowa Space Grant Conference Proceedings, section I-4, 1999. Nutter, W. Forrest, Tylka, G., Basart, J.P., Tim, U.S., "Use of Remote Sensing and GIS Technologies to Diagnose and Quantify Plant Stress in Soybean," Tenth Annual Iowa Space Grant Conference Proceedings, section IV-2, 2000. Nutter, W. F., Jr., Guan, J., and Percifield, R. J., "Assessing crop health in time and space by measuring spectral reflectance of solar radiation from plant canopies," Second International Conference on Geospatial Information in Agriculture and Forestry, 10-12 January, Lake Buena Vista, FL, 2000. Sanchez, Julio, Canton, Maria P. Space Image Processing, CRC Press LLC, 1999. Wang, F., "A fuzzy expert system for remote sensing image analysis," in Proc. IGARSS '89, Vancouver, BC, Canada, vol. 2, pp. 848-851, 1989. Wang, F., "Improving Remote Sensing Image Analysis through Fuzzy Information Representation," Photogrammetric Engineering and Remote Sensing, 56(8):1163-1169, 1990a. Wang, F., "Fuzzy supervised Classification of Remote Sensing Images," IEEE Transactions on Geoscience and Remote Sensing, 28(2):194-201, 1990b. Wang, F., "Integrating GIS's and Remote Sensing Image Analysis Systems by Unifying Knowledge Representation Schemes," IEEE Transactions on Geoscience and Remote Sensing, 29(4): 656-665, 1991.



493