Detection of rice sheath blight for in-season disease management ...

118Z. Qin, M. Zhang / International Journal of Applied Earth Observation and Geoinformation 7 (2005) 115–128field was divided into 11 strips, with a width ofapproximately 40 m for each strip except strip 11,which has a width of approximately 37 m. Field dataof infection severity was estimated at each samplingsite by walking centrally along each strip from east towest. The distance between two adjacent samplingsites was approximately 9 m. Thus for each strip wecollected approximately 50 samples. At each site,several plants were manually examined to estimateaverage infection severity of the site. Three measurementswere used to evaluate the severity: SHBDI =percentage of infected tillers (sheath blight) estimatedby manual counting; SHBDH = height in cm of sheathblight symptoms above ground; TH = height in cm ofplant canopy above soil. Finally, following the methodof Nandakumar et al. (2001), we used the three fieldmeasurements to construct a comprehensive fielddisease index (hereafter field disease index) atlaboratory for general evaluation of infection severityat each site:DI i ¼ SHBDI i SHBDH iTH i; (1)where DI i is field disease index for site i. The firstsampling date (7/13/99) was 10 days past 1/2 in.(1.3 cm) internode elongation. The second samplingdate (7/31/99) was at 100% heading. And the finalrating date (8/26/99) was about a week prior to fielddrain for harvest.3.2. Remote sensing image acquisitionFour airborne remote sensing images wereacquired over the field on July 12, July 24, August4, and August 17, 1999 separately during thesampling period by Positive System Inc., usingADAR (Airborne Data Acquisition and Registration)System 5500. The images had four bands respectivelyin blue: band 1 (450–540 nm), green: band 2(530–600 nm), red: band 3 (610–680 nm) and nearinfrared: band 4 (780–1000 nm) regions. The imageswere stored as 8-bit digital number (DN) valuesranging from 0 to 255. Ground resolution of theimages was 1 m which included approximately 150plants. Fig. 1b shows one image of the field, withbands 3, 2 and 1 as RGB, from which one can clearlysee levees of the rice field. The irregular levees weremainly for farming irrigation management accordingto slope and soil variation.3.3. Image processing to extract dataImage data were extracted to relate to the fieldsampling data. Using software ENVI 3.4, wegenerated subset images of the field from the ADARimages. Eleven strips were divided in each subsetimage to match strip number for field sampling. Thenwe calculated the row and column pixel numbers ofeach strip to decide the image sampling locations ofthe strip and computed approximate coordinate foreach image sampling location. To match fieldsampling, we divided evenly each strip of the subsetimages into 50 spots for data extraction. ENVI spectralfunction was used for data extraction from the subsetimages. To minimize possible bias, we tested 4-pixeland 8-pixel schemes of data extraction for comparison.Fig. 2 shows the comparison results. Littlevariation was observed for the two schemes. T-testvalues for the four bands were 0.555, 0.233, 0.406 and0.287 respectively with d.f. = 32. None of them werestatistically significant at confidence level of 95%.Therefore, we concluded that there was no significantdifference for these two sampling schemes. For thefurther analysis, 8-pixel scheme was used in the studyfor data extraction.3.4. Image index computationThree simple methods were used to capture theextracted image data for the comparison with the fieldFig. 2. Effects of different sampling schemes on sampling resultsfor image 7/24.