Root Rot Detection in Sugar Beet Using Digital Multispectral Video

A study to determine whether root rot infestations can be detected using high spatial resolution digital multispectral video (DMSV) data was conducted in the Imperial Valley of California. DMSV data were collected on two dates over 25 fields during the summer of 1999 using a light aircraft. The DMSV images included four spectral bands centered on 530 nm, 570 nm, 650 nm and 750 nm, with nominal widths of the first two bands being 10 nm and the last two bands 20 nm. Data collection missions were flown on July 15 and July 30 and the fields were harvested in late July and early August. Data from the two narrow band filters (525 535 nm and 565 575 nm) were used to calculate the physiological reflectance index (PRI) while data from the red (640 660 nm) and near infrared (740 760 nm) bands were used to calculate the normalized difference vegetation index (NDVI). The NDVI was found to be a better predictor of percent root rot than the PRI while the change in NDVI between the two flight dates was a better predictor than single date NDVI models. Introduction Root rot is a soil-borne fungal disease that can cause serious damage to sugar beets and reduce yields significantly. Early detection of infestations helps managers prioritize fields for harvest and minimizes production losses. Farmers visually inspect their fields to identify changes in canopy condition (e.g., wilting leaves, dead leaves) that may signify root rot or other types of stress. The large area under sugar beet cultivation in the Imperial Valley of California makes it impractical to implement a comprehensive monitoring program based on ground observations. Furthermore, a regular monitoring schedule is necessary to detect the onset of root rot. High spatial resolution remote sensing with ground resolution elements less than two meters may be a viable technique for the regular monitoring of sugar beet over large areas if spectral bands or spectral vegetation indices (SVIs) can be identified that are responsive to changes in the vegetation canopy which accompany the onset of root rot, and the data can be collected and processed in a timely manner (24 36 hours). A study was conducted in the summer of 1999 to determine whether root rot infestations in the Imperial Valley of California can be detected using high spatial resolution digital multispectral video (DMSV) data. Spectral vegetation indices reduce multispectral data to a single value and have been used widely to infer vegetation characteristics such as leaf area index and biomass (Choudhury, 1988). A common form of a spectral vegetation index based on two spectral bands is: Index = (band 1 band 2)/(band 1 + band 2) (1) The normalized difference vegetation index (NDVI) follows this form with band 1 and band 2 being near infrared and red reflectances, respectively. This index has been shown to have a stable relationship with the fraction of photosynthetically active radiation absorbed by a plant canopy (Prince, 1991). Consequently, the NDVI has also been used to model vegetation growth (Choudhury, 1988). While the NDVI has been used extensively over the past two decades with broad band spectroradiometric data, a new suite of indices have been developed that exploits narrow band spectroradiometric data. The physiological reflectance index (PRI) introduced by Gamon et al. (1992) follows the form of equation 1 and is based on reflectances at 531 nm (band 1) and 570 nm (band 2) wavelengths. Gamon et al. (1997) renamed this the photochemical reflectance index and suggested that the index was a good predictor of photosynthetic radiation use efficiency at the leaf or canopy level, if the canopy is uniform and completely covers the soil background. The index has not shown sensitivity to nutrient stress, but it is affected by water stress with severe wilting (Gamon et al., 1997). Objectives and Approach The goal of this project was to investigate whether root rot in sugar beets could be detected using the NDVI or PRI derived from high spatial resolution digital multispectral video (DMSV) data. The study was also intended to determine whether data collection, processing and analysis could be routinely executed within a 24 36 hour time frame. The turn-around time for collecting data and delivering information to farm managers often limits the utility of remote sensing for agricultural applications. The four-band DMSV camera was fitted with filters having the following nominal bandwidths: 525 535 nm; 565 575 nm; 640 660; nm; 740 760 nm. Two data collection missions were flown in the summer of 1999 over 25 fields in the Imperial Valley of southern California. The first flight was on July 15 and the second on July 30, with all radiometric data being collected in a four-hour window centered on solar noon. The aircraft flew at 1700 m AGL and the DMSV imagery had a ground sampling distance of 1.25 m. Differences between the spectroradiometric properties of light and dark targets that were assumed to be invariant between the two flight dates were used to determine global corrections to normalize the spectral indices for differences in illumination conditions. Detailed data on the spatial distribution and degree of root rot infestation (percent of sugarbeet harvested) were collected in a selected field at the time of harvest (August). The percent root rot was obtained for 13 sub-areas in this field, their areas ranging between 1.2 and 6.5 ha (average = 3.8 ha). In addition to these large area samples, samples based on 20 to 30 sugar beet plants were obtained at six sites in the field prior to harvest. The mean NDVI and PRI values were calculated for the each of the 13 sub areas on the two dates and related to the corresponding percent root rot. The average change in index values between the two dates was also considered a possible predictor variable for root rot. The original 1.25 m DMSV data was resampled to 5 m pixels to minimize misregistration errors for the multidate comparisons. Results and Discussion Although the relationships between spectral vegetation indices and percent root rot are based on only 13 observations, the results do provide insight into the potential of these indices for predicting levels of infestation. Changes in the amount of plant photosynthetic material over time appeared to affect both indices systematically. Green leaf material was decreasing between the two flight dates and the PRI and NDVI values were highly correlated on both dates (flight 1: r = .81; flight 2: r = .90). Therefore, the relationships between percent root rot and the two indices on both flight dates were similar, these relationships being illustrated in Figure 1 and Figure 2. The linear least squares regressions of percent root rot on NDVI and on PRI indicated that the NDVI was a better predictor of root rot infestation than the PRI. The coefficients of determination for the NDVI regressions were 0.397 (flight 1) and 0.420 (flight 2) and the corresponding values for the PRI regressions were 0.256 and 0.394. Both indices were sensitive to the single apparent outlier in the data set (root rot = 12.5%) because of the small sample size (n = 13). Figure 1. Relationship between percent root rot and the PRI for flights 1 and 2 ( PRI rescaled for display). The percent root rot is plotted against the change in NDVI from flight 1 to flight 2 in Figure 3. Areas with the least amount of root rot had the largest decline in NDVI over the two-week period while heavily infested areas had little change in NDVI over this period. From Figure 2, it appears as if the NDVI values for all areas were converging on a common minimum value (soil background/litter value). Since the heavily infested areas were closer to this value at the time of flight 1, the change in NDVI over the two-week period was smaller than the change associated with less infested areas that started with larger NDVI values at flight 1. The change in NDVI in relation to root rot can be viewed another way. The slope of the least squares regression line for the two data sets displayed in Figure 2 becomes steeper over time (flight 1: slope = -143.5; flight 2: slope = -156.1). If earlier missions had been executed, the difference in NDVI values between infested and uninfested sites may have been even greater than that observed in this experiment and the slope of the percent rot NDVI least squares regression line would have been shallower. Future research should examine changes in this relationship over the season to determine the earliest date for predicting root rot infestations. Figure 2. Relationship between percent root rot and the NDVI for flights 1 and 2. Figure 3. Relationship between percent root rot and the change in NDVI (flight 1 minus flight 2). There was a consistent and linear reduction in percent root rot associated with the change in NDVI between flight dates (Figure 3). Two apparent outliers in the data set (root rot = 12.5% and 6.28%) were removed and percent rot was regressed on the change in NDVI. The linear least squares regression model was: Percent rot = -367.0 (NDVI Flight 1 NDVI Flight 2) + 63.9 r 2 = 0.84 (2) We tested this equation using an independent set of root rot observation made just prior to harvest. A sample of 20 30 beets were examined at each of six sites and the NDVI was determined for these locations using the imagery collected on both flight dates. The slightly different location and orientation of the flight frames on the two dates meant that each of the six sites could not be located on both dates. Percent rot was calculated using equation 2 and similar regression equations based on single date NDVI values. The results for the three regression models are presented in Table 1. The percent root rot was estimated using equation 2 for three of the independent plots. The estimated values at plots 5 and 6 were in close agreement with the observed values, while the model estimate of percent rot at plot 2 was higher by a factor of almost two (Table 1). However, plot 2 had the lowest observed infestation (9%) and the lowest value predicted by the model (16%). Predictions based on NDVI values from the individual flights were generally poor for both models (Table 1). Plot ∆NDVI Observed F1 Est. F2 Est. ∆NDVI Est.