Ocean feature extraction from SAR Quicklook Imagery using Convolutional Neural Networks
We used a large dataset of Sentinel-1 and TerraSAR-X quicklook images downloaded from the internet in order to classify the imagery into different classes of subscenes including: open ocean, land, sea ice and ships using Convolutional Neural Network (CNN) classifiers. We construct a training dataset...
| Main Authors: | , , |
|---|---|
| Format: | Conference Object |
| Language: | German |
| Published: |
2018
|
| Subjects: | |
| Online Access: | https://elib.dlr.de/117706/ https://elib.dlr.de/117706/1/hashemi_1570417418.pdf |
| Summary: | We used a large dataset of Sentinel-1 and TerraSAR-X quicklook images downloaded from the internet in order to classify the imagery into different classes of subscenes including: open ocean, land, sea ice and ships using Convolutional Neural Network (CNN) classifiers. We construct a training dataset of subscenes of the images using visual inspection and AIS data. We then focused on the open ocean scenes acquired under different environmental conditions to classify them into different wind speed and sea state categories. We compare the results to wind speed, sea state model results and NOAA buoy measurements. In order to find the subscenes containing ships, icebergs and oil slicks we further utilize the CNN over open ocean and coastal SAR scenes. Statistics on validation is given using categorical cross entropy loss. In addition several high resolution images are used in order to test the performance of the trained Convolutional Neural Network. This study will help to retrieve such images relevant to maritime investigations of ships, oil and environmental parameters using big data methods. |
|---|