A texture-based classification algorithm with histograms of oriented gradients for ALOS/PRISM panchromatic imagery

Panchromatic (PAN) satellite imagery comprises only a single band but it has finer resolution in comparison to the multi-spectral band imagery. In the case of feature extraction and classification, although the multi-spectral imagery has an advantage in availability of the different aspect of spectral properties of the ground coverage, the recent studies introducing intelligent classification and feature extraction increases interest of using object-based classification of PAN imagery, e.g., texture analysis. Wavelet-based method is one of the widely used methods that have been studied in satellite-borne imagery but the classification accuracy is still developing. In this paper, the recent feature extraction method developed in computer vision field, the Histograms of Oriented Gradients (HOG), is newly introduced in classification of satellite-borne PAN imagery. It is tested with the PAN image to evaluate its effectiveness in the case with classification and image recognition of ground objects on PAN image with HOG features. The wavelet-based method, feature extraction with the Gabor filter, is compared with intelligent classifiers, Neural Network and k-Nearest Neighbor algorithm. The test with ALOS/PRISM image demonstrates higher performance of the HOG feature with approximately +10 % increase of over all accuracy, +29 % and +27 % of the producer´s and user´s accuracy at highest and mere -4% decrease of the both accuracy at lowest.