The Effects of Polynomial Interpolation and Resampling Methods in Geometric Correction on the Land-cover Classification Accuracy of Landsat-8 OLI Imagery: A Case Study of Kulon Progo Area, Yogyakarta

摘要

Geometric correction is an important step in image pre-processing, because it determines the the positional accuracy ofthe data. However, the geometric correction also includes pixel values interpolation in their new position, so that it maychange original values. This study objectives were (a) to provide information on the effect of geometric correctionmodels on the accuracy of land-cover classification, especially using per-pixel classification with maximum likelihoodalgorithm; and (b) to assess the effect of image resampling methods on the accuracy of the multispectral classificationresults. This study made use of Landsat 8 OLI Level 1G imagery covering Kulon Progo Area, Yogyakarta, so thatseveral ground control points (GCPs) were needed to suppress geometric errors. Non-systematic geometric correctionwas undertaken using first, second and third order polynomial transformations. After that, several resampling processeswere applied to the geometrically corrected image, i.e. Nearest Neighbour, Bilinear and Cubic Convolutioninterpolations. It was found that the affine transformation using six GCPs distributed over the edges of the image,delivered an RMSE value of 0.355539. In addition, the second order polynomial with 10 GCPs scattered around theedges of the image gave an RMSE value of 0.178053. While the third order polynomial transformation with 17 GCPsthat were evenly distributed in the image produced an RMSE value of 0.100343. The resampling process produced newimages with new pixel values, which were then tested with respect to their classification accuracies based on maximumlikelihood algorithm. Samples for accuracy assessment were taken using stratified random sampling strategy. Sampleswere taken in terms of polygons whose size was determined by considering the pixels’ displacement as the results ofgeometric corrections. This study also found that resampling with nearest neighbour interpolation using third orderpolynomial equation produced the best overall accuracy of 75.46%, with a Kappa of 0.7032.