Vessel Detection From Nighttime Remote Sensing Imagery Based on Deep Learning
Autor: | Ronghao Li, Changyu Luo, Jiangnan Shao, Feixiang Zhang, Yongsheng Zhou, Qingyao Yang |
---|---|
Rok vydání: | 2021 |
Předmět: |
Atmospheric Science
Feature fusion QC801-809 business.industry Computer science Deep learning Geophysics. Cosmic physics nighttime remote sensing Adaptively spatial feature fusion vessel detection Rapid detection Resource protection Ocean engineering YOLOv5 Remote sensing (archaeology) Detection performance Artificial intelligence Computers in Earth Sciences sea–land mask business Scale (map) TC1501-1800 Remote sensing |
Zdroj: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol 14, Pp 12536-12544 (2021) |
ISSN: | 2151-1535 1939-1404 |
DOI: | 10.1109/jstars.2021.3125834 |
Popis: | The continuous and rapid detection of sea vessels on a large scale is of great importance in marine traffic management, resource protection, and rights maintenance. Nighttime remote sensing can reflect human activities during the night with a wide swath and high efficiency, which is unique for vessel detection. Deep learning algorithms have already demonstrated superior performance in many fields, but it is confronted with some problems when applied to vessel detection with nighttime remote sensing imagery, including the lack of labeled dataset, the missed detection of small vessels, and false alarms of land targets. In this article, first, the nighttime remote sensing imagery was collected and the sea vessels in it were labeled. Second, to enhance the detection performance of small vessels, a modified YOLOv5 algorithm—TASFF-YOLOv5 was proposed, which was supplemented with a tiny target detection layer and a four-layer adaptively spatial feature fusion network to obtain a better feature fusion. Third, a land mask operation based on the sea–land prior database was performed to eliminate the false alarms of the land lights. The experimental results showed that the proposed TASFF-YOLOv5 could effectively improve the precision, recall, and mAP0.5 on the vessel dataset, achieving 95.2%, 93.1%, and 94.9% respectively. |
Databáze: | OpenAIRE |
Externí odkaz: |