Abstract:
Machine learning and remote sensing technologies can effectively investigate and monitor the dynamic features of dunes,
sand accumulations, and environmental changes. However, the present article examines the performance of models constructed
using various Linear SVM model implementations during the binary classification task. In R software, three linear SVM libraries
(LibSVM, LibLINEAR, and GLM) are used to find the optimal and accurate spectral sand index. The data of Landsat-8 (OLI)
satellite is used and calibrated to reflectance value, and then 15 sequences of Normalized Difference (ND) are generated. The most
important and weighted NDs among the 15 trained have been chosen (ND34, ND47, and ND57). Nine Linear SVM methods have
been calculated. LibLINEAR library contains several classification types (LibLINEAR1 to LibLINEAR7). The accuracy of the result
images is done by assigning random points to six levels 500 to 1000 points according to the reference image created by supervised
classification. The average Kappa and overall accuracy for all levels of random points show that the optimal three methods are
LibL1, LibL7, and LibL3; with Kappa values of 77.20%, 76.96%, and 76.83%, respectively, and overall accuracy values of 88.60%,
88.48%, and 88.41%, respectively. In contrast, the widely used LibSVM shows less accuracy with more execution time than LibLINEARs.
