Application of deep convolutional neural networks to automatic feature/object detection in high resolution remote sensing imagery

Abstract

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT REQUEST OF AUTHOR.] This thesis investigates how deep learning can be used to automatically detect features and/or objects of interest in high resolution satellite imagery. First an introduction to elementary machine learning algorithms and concepts is provided. This is followed with an overview of neural networks, convolutional neural networks, and a brief presentation of a few modern Deep Convolutional Neural Network (DCNN) architectures. Then we discuss how transfer learning and data augmentation can be used to improve the performance of DCNN and help overcome a shortage of labeled training data. Next, our fusion pipeline and the result of various fusion algorithms on two benchmark remote sensing data sets are presented. Using an evolved Sugeno fuzzy integral we were able to achieve state-of-the-art results on both the UCMerced and RSD benchmark data sets. Next, a world-wide SAM site data set is introduced and we demonstrate the performance of CaffeNet, GoogLeNet, ResNet-50, ResNet-101, on this data set. ResNet-101 achieved a 96.4 percent average accuracy on the world-wide SAM site data set. Finally, we present a China SAM site search and detection case study. First, we compare a visual Broad Area Search (BAS) with a DCNN-assisted (BAS). The DCNN-assisted broad area search found to be 80X faster than the traditional BAS method while yielding the same statistical accuracy (F1 = 90 percent).