Change Detection Of Marine Environments Using Machine Learning
Abstract
Change detection in aerial imagery is important for many disciplines. The Navy and Marine Corps have used it for planning missions in coastal areas. Current change-detection methods analyze registered images pixel by pixel. We trained two convolutional neural networks, VGG19 and DenseNet121, to distinguish eight coastal classes using 8,000 oblique aerial images. We then used both commercial satellite images and orthorectified imagery from small unmanned aerial systems with structure from-motion methods to test whether learning could transfer from the oblique images. Our results showed significant effects of the tile size for coastal classification in the orthorectified images. We also tested change detection by comparing the highest confidence coastal class between images taken in different time periods, but this was often unsuccessful due to the errors in the classifications.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2020
- Accession Number
- AD1114662
Entities
People
- Theodore A. Iii Ayoub
Organizations
- Naval Postgraduate School