We selected training samples from the images figure 3

This preview shows page 7 - 9 out of 27 pages.

We selected training samples from the images (Figure 3) by carefully selecting homogeneous pixels so that every land use/land cover (LULC) class (forest, oil palm, rubber, mangrove, urban, and water bodies) has three sets of training samples (10, 20 and 30 polygons where each polygon contains about 40 60 pixels). The number of pixels (40 or 60) selected for each polygon is dependent on the size of the land use. For example larger number of pixels was selected for oil palm and fewer pixels were used for rubber. Different training samples (10, 20 and 30 polygons) were used to test if MLC and SVM can produce higher accuracy with increased number of training samples. This size of training samples follows the guide where training sample size for each class should be not fewer than 10 30 times the number of bands [40]. We used all the spectral bands of Landsat sensors except for the thermal band for the classification with both MLC and SVM. For SVM we tested all kernel types i.e. , radial, polynomial, linear, and sigmoid [41] and after several trials we chose values of the following parameters that produced the highest accuracies (Bias in Kernel function = 1, Gamma in Kernel function = 0.167, penalty parameter = 100, pyramid level = 0 and class probability threshold = 0). The overall classification accuracies produced by MLC and SVM using 10, 20 and 30 samples were compared using Analysis of Variance (ANOVA). The classification results were validated using another independent set of polygons (10 polygons with 40 to 60 pixels Figure 3) distributed across the study region which we referred to the Johor land
Remote Sens. 2015 , 7 14366 use maps produced by the Department of Agriculture, Peninsular Malaysia (scale 1: 250,000) of 1990, 2000, 2006, 2008 and 2010. Similar to the training dataset we selected more validation pixels for oil palm and fewer pixels for rubber. The Johor land use maps were considered as ground-truth because these maps were produced from aerial photos and SPOT images, and verified by extensive field work. Our local knowledge of several locations also helped us to verify the results. We also used land cover reports produced by the Comprehensive Development Plan ii (unpublished) report produced by the Iskandar Regional Development Authority (IRDA) for years 2013 to 2025. Relative Predictive Error (RPE) was used in this study to quantify the mean percentage difference between land cover classified by digital classification techniques and land cover data produced by the Department of Agriculture (DOA) and IRDA. RPE provides the direction of changes (underestimation or overestimation) in predicted values compared to measured values. Figure 3. The distribution of test samples (30 polygons) and validation samples (10 polygons) for each land cover types in the study area. Yellow color symbols show the test samples and blue color represents the validation samples respectively.

  • Left Quote Icon

    Student Picture

  • Left Quote Icon

    Student Picture

  • Left Quote Icon

    Student Picture