Loading…

Self-Supervised Deep Learning for Vehicle Detection in High-Resolution Satellite Imagery

In this paper, we demonstrate a self-supervised deep learning pipeline that can effectively learn to detect vehicles without using any pre-labeled training data. The pipeline uses a morphological vehicle detection algorithm to automatically generate training sets for a convolutional neural network (...

Full description

Saved in:
Bibliographic Details
Main Authors: Awwad, Zeyad, Alnasser, Faisal, Alshahrani, Tariq, Moraguez, Matthew, Alabdulkareem, Ahmad, De Weck, Olivier
Format: Conference Proceeding
Language:English
Subjects:
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this paper, we demonstrate a self-supervised deep learning pipeline that can effectively learn to detect vehicles without using any pre-labeled training data. The pipeline uses a morphological vehicle detection algorithm to automatically generate training sets for a convolutional neural network (CNN). We tested this methodology on a mixed-use urban neighborhood in Riyadh, Saudi Arabia using 0.31-meter multispectral Worldview-3 satellite imagery with eight bands in the visible and near-infrared wavelengths. This method leverages the class imbalance inherent to many vehicle detection problems by generating a balanced training sample from a high-precision, low-recall morphological model to train a neural network to identify general vehicle characteristics. This approach is built on broadly applicable image processing methods and, with appropriate adjustments, might be adapted to high-resolution from various satellite or aerial sources.
ISSN:2153-7003
DOI:10.1109/IGARSS47720.2021.9554580