*3.1. Experimental Setups*

In this section, experimental analysis about the H2F based classification method are provided. H2F is compared with six recently proposed methods, i.e., Gabor + ELM (GE) [47], LBP + Gabor + ELM (LGE) [46], RGF + Network (RVCANet) [37], RGF + Ensemble (HiFi) [32], and another two methods, edge-preserving filtering (EPF) [58] and intrinsic image (IIDF) [59] based methods. Among these methods, GE and LGE directly concatenate multiple features without further operations. Thus, they could be regarded as the baselines. RVCANet also tries to extract deep features from HSI data, but it adopts a deep network manner. HiFi is a multiple feature fusion method where the results are obtained by weighted voting. These methods have similar motivation as the proposed method, so we use them for comparison. All the methods are compared on two popular (Indian Pines, Kennedy Space Center (KSC) (Available online: http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral\_ Remote\_Sensing\_Scenes)) and one challenging (GRSS\_DFC\_2014 [60,61]) data set. We run the above methods 50 times with randomly selected train and test samples, and the average accuracies and the corresponding standard deviations are reported. Overall accuracy (OA), average accuracy (AA) and kappa coefficient (*κ*) are selected for evaluation [62]. For the three data sets, 20 pixels per class are used for training, and the rests for testing. Some classes (especially in Indian Pines data set) have a total of nearly 20 samples. In this case, we directly use half of them for training and the others for testing. In H2F, we construct nine sub-feature sets (six for LBP, two for Gabor and one for RGF) with nine features per set, totally 81 features. Rolling times of RGF is set as 1–9 with = 1, wavelength *δ* in Gabor is 16, orientation number is 18, and window size in LBP is 3 × 3. Under this setting, Indian Pines/KSC/GRSS\_DFC\_2014 could be represented by 225792/198144/92160 dimensional features, but only 1% around are non-zero. The hyper-parameters in ELM (linear kernel) is determined by five-fold cross validation. The regularization coefficient is chosen from a set {1, 10, 100, 1000}, and the hidden neuron number is chosen from {100, 200, ··· , 2000}. According to the results of cross validation, 1000 and 100 are appropriate for the above two parameters.

## *3.2. Data Sets*

• Indian Pines: This data is widely used in HSI classification, which was gathered by airborne visible/infrared imaging spectrometer (AVIRIS) in Northwestern Indiana. It covers the wavelengths ranges from 0.4 to 2.5 μm with 20 m spatial resolution. In total, 145 × 145 pixels are included and 10,249 of them are labeled. The labeled pixels are classified into 16 classes. There are 200 bands available after removing the water absorption channels. A false color composite image (R-G-B=band 36-17-11) and the corresponding ground truth are shown in Figure 3a,b.


**Figure 3.** *Cont.*

**Figure 3.** False color composite images of (**a**) Indian Pines; (**c**) KSC and (**e**) GRSS\_DFC\_2014 data sets and the ground truths (**b**,**d**,**f**). Each color corresponds to a certain class.
