Object Detection in Satellite Images Using Modified Pyramid Scene Parsing Networks

This chapter aims at developing a deep neural network using PSPNet architecture with modifications for detecting specified objects in satellite images provided to the Kaggle competitors. It consists of the major steps, including the adaptation of convolutional neural networks to multispectral image data and evaluation of data fusion strategies for semantic segmentation on satellite images and introduction of a combined training objective for defining the desired output for the purpose of image segmentation. For evaluating the classification results, two metrics such as pixel accuracy and Jaccard index are used. The proposed approach has successfully handled the problems, which significantly improves the quality of the final models. It incorporates several steps such as the adaptation of fully convolutional networks to multispectral satellite images and the evaluation of several data fusion strategies on semantic segmentation tasks of satellite images with a combined training objective.