Optics and Precision Engineering, Volume. 31, Issue 13, 1962(2023)
MMShip： medium resolution multispectral satellite imagery ship dataset
Figures & Tables(13)
Fig. 1. Processing flows of medium resolution multispectral satellite imagery ship dataset
Table 1. Comparison of remote sensing data sets containing ships
View tableView in ArticleTable 1. Comparison of remote sensing data sets containing ships
Datasets Source Number Pixel Class Resolution/m Year NWPUVHR-10 Google Earth 800 1k×1k 10 0.5-2 2014 LEVIR［ 16 ］Google Earth 22k 800×600 3 0.2-1 2017 DOTA Google Earth/GF-2/JL-1 2 806 800×800-4 000×4 000 15 ≤0.5 2017 NWPU-RESISC45［ 17 ］Google Earth 31 500 256×256 45 0.2-30 2017 HRRSD［ 18 ］Google Earth/Baidu Map 21 761 —— 13 0.15-1.2 2018 DIOR Google Earth 23 463 800×800 20 0.5-30 2018 xView［ 19 ］WorldView 3 1 412 1 500×1 500 60 ≤0.3 2018
Table 2. Comparison of remote sensing ship data sets
View tableView in ArticleTable 2. Comparison of remote sensing ship data sets
Datasets Source Number Pixel Class Resolution/m Band Year HRSC2016 Google Earth 1 070 300×300-1 500×900 25 0.4-2 RGB 2016 Kaggle［ 20 ］SPOT6/7，Pléiades 192 556 768×768 1 0.3-6 RGB 2018 MASATI［ 21 ］Microsoft Bing maps 6 212 512×512 1 —— RGB 2018 FGSD Google Earth 2 612 930×930 43 0.12~1.93 RGB 2020 ShipRSImageNet Multiple 3 435 930×930-1 500×1 500 50 0.12~6 RGB 2021 FGSC-23［ 22 ］Multiple 4 052 —— 23 0.4~2 RGB 2021 MMShip Sentinel-2 5 016 512×512
10 980×10 980
1 10 RGB/NIR 本文
Table 3. Configuration information
View tableView in ArticleTable 3. Configuration information
Configuration Version Python 3.7.11 Pytorch 1.8.0 Torchvision 0.9.0 MMdetection 2.22.0 mmcv 1.4.6 CUDA 11.1
Table 4. Comparison of detection algorithms on test sets
View tableView in ArticleTable 4. Comparison of detection algorithms on test sets
Method Backbone NIR（%） RGB（%） Multispectral（%） AP50 AP75 AP90 AP50 AP75 AP90 AP50 AP75 AP90 One-stage Anchor-based SSD512 VGG16 74.11 44.92 1.47 75.39 41.67 3.03 74.62 44.16 4.89 One-stage Anchor-based RetinaNet ResNet50 88.32 46.95 3.03 87.39 38.90 2.27 67.50 41.86 3.41 One-stage Anchor-free FoveaBox ResNet50 82.51 45.17 9.09 76.97 36.80 9.09 79.17 41.19 7.03 One-stage Anchor-free FCOS ResNet50 89.12 47.68 0.71 88.90 46.85 2.27 88.99 46.83 1.52 Two-stage Faster R-CNN ResNet50 77.41 50.88 2.61 87.46 49.40 0.73 82.41 52.77 2.14 Two-stage Cascade R-CNN ResNet50 90.39 53.31 1.30 90.46 52.72 3.03 90.57 52.13 9.09
Table 5. Learning rate setting
View tableView in ArticleTable 5. Learning rate setting
Method lr SSD512 0.000 5 RetinaNet 0.010 0 FoveaBox 0.005 0 FCOS 0.010 0 Faster R-CNN 0.020 0 Cascade R-CNN 0.020 0
Table 6. HRSC2016 test results
View tableView in ArticleTable 6. HRSC2016 test results
Method Backbone RGB/% AP50 AP50 AP75 Rotated-RetinaNet ResNet50 84.263 59.07 10.48 Rotated-FCOS ResNet50 88.52 65.76 7.93 Rotated-Faster R-CNN ResNet50 78.99 43.91 9.09
Tools
Get Citation
Copy Citation Text
Li CHEN, Linhan LI, Shiyong WANG, Sili GAO, Xiangzhou YE. MMShip： medium resolution multispectral satellite imagery ship dataset[J]. Optics and Precision Engineering, 2023, 31(13): 1962
Paper Information
Category: Information Sciences
Received: Sep. 6, 2022
Accepted: --
Published Online: Jul. 26, 2023
The Author Email: WANG Shiyong (s_y_w@sina.com), GAO Sili (s_y_w@sina.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20