Rotation invariant binary gradient contour for geographic object-based image analysis

Abstract

This study proposed a modified rotation invariant texture descriptor based on Binary Gradient Contour (BGC1) for land cover classification under Geographic Object-based Image Analysis (GEOBIA). The modified texture descriptor’s performance was tested with 6 machine learning algorithms and a high-resolution Theos satellite image of the area of the city of Khon Kaen. The satellite image was segmented into 9,929 homogeneous land cover objects, of which, 5,417 objects were labeled as one of the ten land cover classes and validated using the 5-Fold cross validation method. The overall accuracy, the individual class F1-Scores, and the computational efficiency of the classification models, which used rotation invariant BGC1_Rot, were compared with models, which had used GLCM, LBP variations, and the original BGC1. The results showed that among the 6 classifiers, Random Forest (RF) had produced the best overall accuracy. The model with RF and BGC1_Rot had produced the best overall accuracy at 84.863%, which was significantly higher than the original BGC1, and was the highest F1-score for 6 out of 10 investigated land cover classes. During the feature extraction step, the more computationally efficient BGC1_Rot was also found to process 4.48 times faster than GLCM. When compared to the widely accepted Uniform LBP_Uni, BGC1_Rot provided an overall accuracy and an average F1-Score that were slightly better with a similar computation time. Thus, the proposed BGC1_Rot has been proven to be an effective texture descriptor for GEOBIA based on overall and individual class accuracy, as well as on computational efficiency.