MODIFICATION OF INPUT IMAGES FOR IMPROVING THE ACCURACY OF RICE FIELD CLASSIFICATION USING MODIS DATA

Abstract

The standard image classification method typically uses multispectral imageryon one acquisition date as an input for classification. Rice fields exhibit high variability inland cover states, which influences their reflectance. Using the existing standard method forrice field classification may increase errors of commission and omission, thereby reducingclassification accuracy. This study utilised temporal variance in a vegetation index as amodified input image for rice field classification. The results showed that classification ofrice fields using modified input images provided a better result. Using the modifiedclassification input improved the correspondence between rice field area obtained from theclassification result and reference data (R2 increased from 0.2557 to 0.9656 for regencylevelcomparisons and from 0.5045 to 0.8698 for district-level comparisons). Theclassification accuracy and the estimated Kappa value also increased when using themodified classification input compared to the standard method, from 66.33 to 83.73 andfrom 0.49 to 0.77, respectively. The commission error, omission error, and Kappa variancedecreased from 68.11 to 42.36, 28.48 to 27.97, and 0.00159 to 0.00039, respectively, whenusing modified input images compared to the standard method. The Kappa analysisconcluded that there are significant differences between the procedure developed in thisstudy and the standard method for rice field classification. Consequently, the modifiedclassification method developed here is significant improvement over the standardprocedure.