For the Gray-Potsdam data set, the thresholds obtained with the previous formula were not sufficient to cover the size variations of the targeted classes. To obtain a set of thresholds spanning a larger range of values without increasing the number of thresholds, we considered the 14th first values of the geometric sequence whose nth term is given TABLE 4. THE EXPERIMENTAL SETTINGS EMPLOYED IN THE REPORTED TESTS. TABLES WITH EVALUATION RESULTS Table 6 Table 8 Table 10 Table 11 Table 12 DATA SET SPLIT Pavia1 GrayPotsdam1 GrayPotsdam1 Pavia 1 Pavia1 ATTRIBUTES All All All All All TREE TYPES All All All All All QUANTIZATION (B) 8 8 8 8 8 Table 13 Table 14 Table 15 Table 16 Table 17 Table 18 Pavia1 Pavia2 Pavia2 GrayPotsdam1 GrayPotsdam2 Potsdam2 Potsdam3 Potsdam3 All All All All All All All All All All All All 16 32 8 8 8 8 8 8 CONNECTIVITY POSTPROCESSING THRESHOLDS All 4 4 4 8 4 8 4 All FPs LFAPs FPs LFAPs All All Manual Without Manual Without All All 4 FPs 4 4 4 4 4 FPs FPs FPs FPs FPs All All Manual Without Manual Without Manual Manual TABLE 5. THE PAVIA AND GRAY-POTSDAM DATA SET SPLITS EMPLOYED IN THE REPORTED TESTS. DATA SET SPLIT DESCRIPTION Pavia1 (Figure 3) Pavia2 [Figure 8(b)] ◗ Standard split obtained from http://dase.grss-ieee .org/ ◗ Commonly used in the literature ◗ 10 sets of training/test samples randomly extracted from a restricted region of the Pavia data ◗ Lower variability of training pixels when compared to Pavia1 ◗ Better separation among training/testing pixels than Pavia1 ◗ The same number of training samples per class as Pavia1 GrayPotsdam1 ◗ 10 sets of training/test samples randomly extracted from all of the ground-truth CCs of Gray-Potsdam ◗ The same number of training samples per semantic class GrayPotsdam2 (Figure 9) GrayPotsdam3 (Figure 11) ◗ 10 sets of training/test samples randomly extracted from a restricted set of ground-truth CCs of Gray-Potsdam ◗ The same number of training and test samples as GrayPotsdam1 ◗ Data set divided in half ◗ The same number of training pixels per class as GrayPotsdam1 and GrayPotsdam2, but a different number of test pixels ◗ Test set composed of all pixels from the lower half ◗ 10 sets of training samples randomly extracted from the upper half SEPTEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 53 ◗ Similar to Pavia2: introduce CCs in the test set that do not include any training pixels. ◗ Test if the features extracted from APs really reflect the geometrical characteristics of the objects belonging to a certain class or if the success of APs is mainly due to the leakage of training/testing features ◗ Generalize APs to multiple-image data sets ◗ Compute training and testing features from different trees obtained from the two halves of the image, with no leakage of training/ testing features OBJECTIVE ◗ Provide evaluation scores of AP extensions, which are comparable with the results in the literature (in which all the whole data are preprocessed with PCA, and training/testing features are extracted from the same tree, leading to the leakage of training features) ◗ Evaluate the impact of having large ground-truth regions not contributing to the training set ◗ Generalize APs to data sets with lower levels of leakage of training features, which still remain due to all the whole data being preprocessed with PCA, as well as to training and testing features being extracted from the same tree ◗ Provide a more realistic partition of the Pavia University data set when compared to Pavia1 ◗ Evaluate APs and their extensions on Gray-Potsdam using the training/test splitting method commonly used in the literaturehttp://dase.grss-ieee.org http://dase.grss-ieee.org