Tree-based convolutional neural networks for object classification in segmented satellite images

Abstract

Satellite images have a very high resolution, which make their automatic processing computationally costly, and they suffer from artifacts making their processing difficult. This paper describes a method for the effective semantic segmentation of satellite images, and compares different object classifiers in terms of accuracy and execution time. In the paper, the image spectrum is used to reduce the computational cost during the segmentation and classification steps. Firstly, artifacts are corrected from the satellite images for facilitating the feature extraction process. After this, semantic representation is used to gather the semantic regions of downscaled images. As the images are very large, this scaling down significantly reduces the computing time with little degradation in the coarse object detection results. A deep neural forest classifier finds potential regions before executing the pixel-based segmentation. Finally, in our experiments, boundary detection and several classifiers are evaluated to find the objects associated with these regions. The paper details the set-up for our tree-based convolutional neural network. The results indicate that this tree-based convolutional neural network outperforms the other surveyed techniques in the literature.