Self-adaptive Image Segmentation Optimization for Hierarchal Object-based Classification of Drone-based Images

Abstract

This study proposes an approach for the quality improvement of feature extraction in unmanned aerial vehicle (UAV)-based images through object-based image analysis (OBIA). A fixed-wing UAV system equipped with an optical (red–green–blue) camera was used to capture very high spatial resolution images over urban and agricultural areas in an arid environment. A self-adaptive image segmentation optimization aided by an orthogonal array from the experimental design was used to optimize and systematically evaluate how OBIA classification results are affected by different settings of image segmentation parameters, feature selection, and single and multiscale feature extraction approaches. The first phase encompassed data acquisition and preparation, which included the planning of the flight mission, data capturing, orthorectification, mosaicking, and derivation of a digital surface model. In the second phase, 25 settings of multiresolution image segmentation (MRS) parameters, namely, scale, shape, and compactness, were suggested through the adoption of an L25 orthogonal array. In the third phase, the correlation-based feature selection technique was used in each experiment to select the most significant features from a set of computed spectral, geometrical, and textural features. In the fourth phase, the ensemble adaptive boosting algorithm (AdaBoost) was used to classify the image objects of segmentation levels in the orthogonal array. The overall accuracy measure (OA) and kappa coefficient (K) were computed to represent a quality indicator of each experiment. The OA and K values ranged from 89% to 95%, whereas the K values ranged from 0.75 to 0.95. The MRS parameter settings that provided the highest classification results (>94%) were analyzed, and class-specific accuracy measures and F-measure were computed. Multiscale AdaBoost classification was conducted on the basis of the computed F-measure values. Results of the multiscale AdaBoost classification demonstrated an improvement in OA, K, and F-measure.