Abstract

Processing of satellite images using deep learning and computer vision methods is needed for urban planning, crop assessments, disaster management, and rescue and recovery operations. Deep learning methods which are trained on ground-based imagery do not translate well to satellite imagery. In this thesis, we focus on the tasks of semantic segmentation and change detection in satellite imagery. A segmentation framework is presented based on existing waterfall-based modules. The proposed framework, called PyramidWASP, or PyWASP for short, can be used with two modules. PyWASP with the Waterfall Atrous Spatial Pooling (WASP) module investigates the effects of adding a feature pyramid network (FPN) to WASP. PyWASP with the improved WASP module (WASPv2) determines the effects of adding pyramid features to WASPv2. The pyramid features incorporate multi-scale feature representation into the network. This is useful for high-resolution satellite images, as they are known for having objects of varying scales. The two networks are tested on two datasets containing satellite images and one dataset containing ground-based images. The change detection method identifies building differences in registered satellite images of areas that have gone through drastic changes due to natural disasters. The proposed method is called Siamese Vision Transformers for Change Detection or SiamViT-CD for short. Vision transformers have been gaining popularity recently as they learn features well by using positional embedding information and a self-attention module. In this method, the Siamese branches, containing vision transformers with shared weights and parameters, accept a pair of satellite images and generate embedded patch-wise transformer features. These features are then processed by a classifier for patch-level change detection. The classifier predictions are further processed to generate change maps and the final predicted mask contains damage levels for all the buildings in the image. The robustness of the method is also tested by adding weather-related disturbances to satellite images.

Library of Congress Subject Headings

Remote-sensing images--Data processing; Remote-sensing images--Classification; Deep learning (Machine learning); Computer vision

Publication Date

10-2021

Document Type

Thesis

Student Type

Graduate

Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)

Advisor

Andreas Savakis

Advisor/Committee Member

Alexander Loui

Advisor/Committee Member

Guoyu Lu

Campus

RIT – Main Campus

Plan Codes

CMPE-MS

Download

Share

COinS