Using Deep Convolutional Neural Networks to Classify Littoral Areas With 3-Band and 5-Band Imagery
Abstract
With the assistance of high-resolution satellites, unmanned aerial vehicles, and fixed camera observation points, coastal change detection and landscape classification are active research areas that have the capability to provide situational awareness. However, classification of bottom types in littoral waters is an area of coastal landscape classification that has not been studied extensively, and accurate and timely classification of bottom types remains elusive. Furthermore, it is unclear whether 5-band imagery (RGB, or red, green, blue; along with near infrared and RedEdge) will help deep convolutional neural networks (DCNN) classify bottom types easier than just color (RGB). In this study, a DJI Inspire unmanned aerial vehicle equipped with a MicaSense RedEdge sensor was used to obtain 5-band imagery of several coastal areas. These images were classified by various means for six areas: swash zone, sandy bottom, bottom other than sand, sand, kelp and above ground rock. This database was then used to train the DCNN for classification on unseen imagery. The models were first initialized with RGB data and then compared to the 5-band outputs. DCNNs were able to classify littoral areas with more accuracy using 5-band imagery than 3-band imagery. Further studies can apply the methods developed in this research and compare 5-band imagery obtained from unmanned aerial systems with imagery obtained from high-resolution satellites such as WorldView 3.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2020
- Accession Number
- AD1114260
Entities
People
- Ashley M. Mielke
Organizations
- Naval Postgraduate School