An integrated methodology to improve classification accuracy of remote sensing data

In this paper, we investigated and improved the accuracy of supervised classification by eliminating electromagnetic radiation scattering effect of aerosol particles from cloud-free Landsat TM data. The scattering effect was eliminated by deriving a mathematical model including the amount of the scattered radiation per pixel area and aerosol size distribution, which was derived using randomly collected training sets. An algorithm in C++ has been developed with iterations to derive the aerosol size distribution and to remove the effect of aerosols scattering in addition to the use of IRDAS software (commercial software). To assess the accuracy of the supervised classification, results of remote sensing data were compared with Global Positioning System (GPS) ground truth reference data in error matrices (output results of classification). The results of the corrected images show great improvement of image quality and classification accuracy. The misclassified off-diagonal pixels were minimized in the accuracy assessment error matrices. Therefore it fulfills the criteria of accuracy improvement. The overall accuracy of the supervised classification is improved (between 18% and 27%). The Z-score shows significant difference between the corrected data and the raw data (between 4.0 and 11.91) by employing KHAT statistics evaluation.