Open-Source Assessments of AI Capabilities: The Proliferation of AI Analysis Tools, Replicating Competitor Models, and the Zhousidun Dataset (2024)

1 Uncovering the Zhousidun Database

In August of 2023, researchers at the Berkeley Risk and Security Lab came across a dataset of Chinese origin named Zhousidun (or “Zeus’s Shield,” using the Chinese colloquial name for the U.S. Aegis Combat System).¹¹1The full Mandarin name for the Aegis Combat System is Zhousidun Zhandou Xitong. The dataset was made publicly available on Roboflow, a platform where users can share datasets and train machine learning models, likely by ShanghaiTech University.²²2We believe that this dataset is most likely from ShanghaiTech’s Visual and Data Intelligence Center (VDI Center). The name of the account that posted the dataset to RoboFlow was “shanghaitech-faxfj.” The Zhousidun dataset is composed of 608 oblique and satellite images of American Arleigh Burke-class destroyers and other allied destroyers and frigates (a complete list of which is provided in Appendix 6).

The nature of this dataset is highly specific and atypical — academic work in this space is largely focused on whole ship detection, not on specific components of military vessels. In those Zhousidun images with Arleigh Burke-class destroyers, the ships’ radar systems (which are part of the Aegis Combat System³³3An integrated, networked combat system deployed on U.S. Navy destroyers and cruisers for fleet air and missile defense.), are labeled with bounding boxes on each image (see Figure 1). The images in the dataset are obtained from publicly available sources with some watermarked with media outlet information and the Google Earth logo. Due to the targeted, military nature of the dataset and the likely academic origins of the account sharing it, we suggest that it is likely that this dataset was accidentally published.

In this paper, we demonstrate a quantitative method to assess the AI capabilities of a third-party through the replication of a near-state-of-the-art computer vision model on the Zhousidun dataset. This method can be applied to other, similar, databases to further our understanding of competitor models and AI capabilities. We then demonstrate the performance of the trained model on this exposed dataset. Furthermore, we create a synthetic data generation process around a U.S. destroyer from the perspective of a small unnamed aerial system and demonstrate the performance of the trained model on this data. Finally, we hypothesize as to why Chinese affiliates may be using this approach as opposed to others and its potential implications for the U.S. Navy.

2 The Proliferation and Lifecycle of AI Intelligence Tools

While the dataset is focused on warships, any conclusions about its intended use must be caveated by the limited information we have about its origins. We cannot say with certainty whether this database was developed by a researcher at a university interested in testing new tools for ship detection, or a navy computer scientist as part of a military project, or even an undergraduate student looking to impress potential employers in the defense sector. That all three of those profiles could fit as creators of this database is a testament to the increasing availability of AI intelligence tools outside of traditional military domains and the reduced skill barrier to using them. In fact, in this scenario, it cannot be assumed that the military scientist would even be the most capable in compiling or training models. Based on the username of the person who released the dataset, it would appear the Zhousidun originated from ShanghaiTech University. We can deduce no additional information about the origin of the dataset.

While historically, military R&D has kept up with, or even led, innovation in the private sector, this is no longer the case. Exquisite military technologies were generally developed in isolation and kept secret, with the know-how and talent to develop these technologies further cloistered, ensuring that the replication of the capability was near impossible externally. The world of artificial intelligence, however, represents a marked exception to this status quo, as AI research and development has been widely democratized. Pioneering artificial intelligence research is, largely, not kept for private profit but is instead almost immediately released into the public domain via publication and paired release of code. The tools, techniques, and procedures to replicate and extend these capabilities are available at the click of a button and the skills required to use them are learnable through the course of a day on YouTube.⁴⁴4You can see excellent courses such as Practical Deep Learning for Coders or Stanford’s CS229 lectures.

With the realization that academics, companies, and governments alike are all using the same methods, and often the same publicly released code, replicating the models that are utilized by adversaries becomes almost trivial. Indeed, the differentiating factor between capabilities is increasingly the data used to train the models as well as the pipelines surrounding data acquisition, model verification and validation, and testing shankarOperationalizingMachineLearning2022 .

Data and methods to source said data have become closely guarded secrets by almost all organizations in the pursuit of “better” AI capabilities. It is rare to find data used by organizations for proprietary or restricted AI capabilities online, making replication of these models difficult. Even “open” AI models, those that publicly release their model weights such as Meta’s Llama or Google’s Gemma, do not release the underlying data that they were trained upon.⁵⁵5With the exception of a few new models, such as the Allen Institute’s OLMo model.

When datasets from adversaries or competitors are accidentally revealed, they offer an opportunity for deep, targeted exploration of the capabilities of models trained on those datasets and the limits of the collection capabilities of those adversaries. Additionally, advances in simulation and modeling make it possible to meaningfully replicate data, thus providing a more holistic testing distribution than the original dataset alone. In cases where no dataset is available, simulation and modeling make it possible to bootstrap a testing dataset entirely.

2.2 AI Ship Detection

One area where academic computer vision research has overwhelmingly embraced expanded public access to satellite imagery is that of ship detection and identification. Researchers are able to identify and track objects and vessels at sea by applying computer vision algorithms to satellite imagery.

These techniques have both military and civilian value. On the civilian side, being able to monitor and identify objects at sea from commercial satellite imagery can enable tracking illicit fishing, identifying ships for sanction evasion, and monitoring climate and environmental changes such as tracking coral reefs degradation. Of course, these very techniques can also be used to identify and track, and even target, naval vessels and other at-sea military equipment.

To highlight this emerging area of research, and also identify ways in which researchers (in government, in academia, in the military, or elsewhere) may approach a similar dataset, we conducted a literature review of 76 papers related to AI ship detection or similar research. A complete bibliography is provided in Appendix 7.

Collectively, these papers demonstrate a substantial interest in high-fidelity maritime object detection. Many of these papers incorporate statistical and computer vision-based techniques to detect and classify ships against crowded backgrounds, including in ports and along coastlines. Primarily, synthetic aperture radar (SAR) satellite imagery serves as the foundation for AI-based ship detection literature, followed by high-resolution electro-optical satellite and oblique imagery. Researchers often retrieve imagery from publicly available sources for most datasets, including the European Sentinel-1 SAR constellation and imagery scraped from Google Earth. In some instances, datasets use imagery from hard-to-obtain civilian sources, like the Chinese Gaofen constellation, while others depend on maritime surveillance feeds from drones and cameras.

The detection of vessels, especially ones aiming to stay hidden on a vast ocean, remains an unsolved and complicated problem. As outlined in our analysis of datasets above, almost all modern ship detection methods attempt to detect entire classes of ships at once. That is, the entire ship has a bounding box drawn around it. This whole-ship detection and classification approach results in models that are unable to properly differentiate between different types of ships. Often this is due to a limitation in deep learning methods to learn salient features that distinguish a specific kind of ship.

3 Replicating Models on Zhousidun

Zhousidun provided the opportunity for us to demonstrate what we believe is a useful approach to testing and assessing competitor datasets, and by extension, models. In sum, we utilized a widely-used, near-state-of-the-art object detection model to analyze the effectiveness of a model trained on Zhousidun on the real-world task of detecting Aegis combat systems in the wild. We perform external validation of this model on synthetic data generated from a custom-built 3D scene of an Arleigh Burke-class destroyer. We then demonstrate the limitations of the Zhousidun dataset, and therefore, of datasets scraped from the public internet, and discuss how such models might be used in the context of Zhousidun’s focus on U.S. and allied naval vessels.

3.1 Data

The Zhousidun dataset is composed of 608 images of military naval vessels, largely destroyers, that are equipped with the Aegis combat system. These images are from ground-based, oblique optical sensors and optical satellite imagery. Bounding boxes have been drawn around SPY radars on the superstructure, one on port and one on starboard, as well as around the vertical launching systems towards the bow and towards the stern of the ship.

The images comprising Zhousidun appear to have been collected from the open Internet due to the inclusion of various watermarks in the images, as well as the unconstrained quality and geometries of the collected images themselves. Indeed, we were able to locate the original sources for the majority of the oblique images in the dataset from a variety of media and military outlets. The satellite images are all low-quality screenshots from Google Earth. Scraping satellite imagery from Google Earth is a common practice in widely-used Chinese academic datasets such as MLRSNet qiMLRSNetMultilabelHigh2020 , AID xiaAIDBenchmarkData2017 , WHU-RS19 daiSatelliteImageClassification2011 , and others.

Together, this means that Zhousidun is not a representative dataset of the sensors and images collected by the PLAN. However, bootstrapping powerful object detectors from low-quality, web-scraped imagery is effective and a popular approach in the machine learning community with the advent of webly supervised chenWeblySupervisedLearning2015 and self-supervised learning techniques.

To ensure that our analysis is representative of realistic, high-quality data of the kind that can be obtained from modern optical sensors, we build a synthetic scene of a static Arleigh Burke-class destroyer in the middle of an ocean in moderate sea conditions (Beaufort force 4). Images are collected in a half-dome around the vessel representative of imagery captured from a small, unmanned aerial system or a high-resolution optical satellite. These images are solely used for evaluation.

Appendix 5 contains further details about how the data is processed for use in a machine learning pipeline.

3.3 Methodology

To conduct our analysis, we use the train and test sets of the Zhousidun dataset as provided by the dataset uploaders. We train the YOLOv8-large (YOLOv8l) detector on the train split of the dataset.¹²¹²12Code and weights are made available on BRSL’s GitHub repo. Further details about the model configuration are provided in Appendix 5.

In this evaluation, we aimed to understand how well a model trained on publicly-available imagery can perform in real-world identification targeting situations. We evaluate the performance of the model through a metric called “mean average precision” (mAP). The mAP metric tells you how “good” a model is at not only finding all of the right objects (a metric called “recall”) but also ensuring that it only gets the right objects (measured through “precision”). mAP ranges between 0 and 1; a higher mAP score indicates that the model is more accurate and misses few objects. We evaluate mAP an intersection over union (IoU) of 0.50, i.e., a predicted bounding box is only correct if it overlaps at least 50% of the true bounding box.

Evaluated simply on the test set of Zhousidun, the trained model achieves a mAP (at 0.50 intersection over union (IoU)) of 0.926. In many targeting workflows, this mAP is considered to be quite good. These results are visualized in Figure 5.

However, the test set of Zhousidun is not a realistic testing target. It represents an artificial set of images collected from an unrealistic source such as Google Images and does not mirror what the real-world performance of a model trained on this publicly-sourced data would look like.

To assess the out-of-distribution performance of this trained model, we generated a synthetic 3D scene of an Arleigh Burke-class destroyer on the open ocean. We utilize the same evaluation methodology as above for these synthetic images. Specifically, we evaluated the performance of the model on all geometries of collect, a subset of collects from oblique viewing angles, and a subset of collects from near-nadir angles mimicking satellite imagery. Outputs from this experiment are provided in Figure 6.

Across all geometries, the model had a mAP of 0.411 with fairly poor recall (0.25) and decent precision (0.85). This means that the model struggled to initially find the components of the Aegis combat system, but when a component was identified, the resulting detection was precise. When focusing on just oblique geometries, the model’s mAP improved to 0.51 with better recall (0.33) and precision (0.87). However, the model performed extremely poorly on geometries mimicking that of a satellite, achieving only a mAP of 0.18 with a recall of 0.20 and a drastically lower precision at 0.18.

Overall, a model trained on Zhousidun has limited targeting capabilities in the real world. It is unlikely that any military would field a model with these performance characteristics. However, it is extremely interesting that training on a small set of unconstrained, publicly available imagery offers such a great starting point to building a robust targeting model. Collecting real images of Allied destroyers with sUAS is not a trivial task and the amount of available data is low. By combining datasets like Zhousidun with a small set of real-world images in tandem with modern approaches in pre-training and fine-tuning, we can convert this weak model into a powerful detector.

4 Why This Matters: Adaptive Net Assessment and the Open-Source

In addition to the specifics of the Zhoushidun dataset leak, this episode underscores the potential incorporation of AI into net assessment. Machine learning capabilities are poised to become a core component of many military programs in the near future—particularly as it relates to intelligence, surveillance, and reconnaissance. Understanding how these capabilities work and how well they work represents a central part of the net assessment capabilities for any nation.

With the growing significance of quantitative capabilities, net assessment must integrate new quantitative methods for model evaluation. We can reconstruct theoretical workflows for analysis by utilizing knowledge of existing data and deployment workflows. While qualitative insights into bureaucratic politics and command control remain crucial, incorporating quantitative techniques enhances our understanding of how technical constraints impact the military balance.

We recommend adopting and scaling the simulation and mirroring approach to the assessment of machine learning models exemplified here. State-of-the-art models are available in the open source and can be obtained and executed cheaply as a white box. Input data to train these models can be created with high fidelity using simulation and modeling tools or collected from the open source, as in this case. Critically, the entire pipeline can be modified at will to understand how varying inputs affect the veracity and reliability of the models being assessed. OSINT analysts can be deployed at many agencies, governmental or non-governmental, to unearth datasets that might be used to train machine learning models to bootstrap the net assessment of a given capability.

Open-source intelligence can provide important insight into how public academic techniques are evolving, as well as any public techniques from the private sector. This is a critical insight as developments in the academic and commercial AI sector are a harbinger for military AI capabilities.Here, we demonstrate that an analysis of the Zhoushidun dataset as a quantitative assessment capability represents a potentially useful predictor of China’s machine-learning arsenal. We hope that replication of this method using similar candidate datasets might present new opportunities for the future net assessment of AI models and use cases. Repeated use of such a methodology on new datasets will provide tighter bounds and more accurate assessment of competitor AI capabilities.

Acknowledgements

We would like to acknowledge the support of the Founder’s Pledge Fund for their support of the Berkeley Risk and Security Laboratory and this work. We thank Trevor Darrell, Suzanne Petryk, and other colleagues for their comments and feedback on this publication.

{appendices}

5 Appendix A: Model Configuration

The Zhousidun dataset is split into train and validation sets of 557 and 51 images respectively. For the purposes of this work, the validation set is used as an external test set and is not used for training or hyperparameter tuning.

The original images are all of varying resolutions. However, a limitation of the YOLO-family of architectures is that input images must all be of the same size. Therefore, each image is resized to be a square 640 x 640px resolution by the Roboflow platform, as visualized in Figure 7. This may have negative consequences on the performance of the model as the aspect ratio of the Aegis components is sometimes drastically changed, but is otherwise a common practice when using CNN-based object detectors.

The YOLOv8l model used in this work is trained for 200 epochs using the AdamW optimizer. The Ultralytics implementation utilized a heuristic to set an optimal learning rate for the optimizer, resulting in a learning rate of $\alpha$=7.14e-4 and momentum $\beta$=0.9. We utilize parameter group weight decay with parameter groups, set automatically by the Ultralytics library:

The training is accelerated with 10 NVIDIA P100s with a batch size of 200. The gain in precision, recall, and mAP plateaus approximately at epoch 150. We continue the training for the full 200 epochs for completeness but report the highest mAP achieved during the training run. Loss curves for the training run are provided in Figure 8.

6 Appendix B: Ships Identified in Zhousidun

We created a list of all identifiable ships in the Zhousidun dataset using the hull classification numbers visible in the images. Many images had unreadable hull classification numbers either due to poor image quality or capture angle. Therefore, this list may not be complete.

7 Appendix C: Papers Reviewed

References

• [1]Airbus Ship Detection Challenge.https://kaggle.com/competitions/airbus-ship-detection.
• [2]Ships in Satellite Imagery.https://www.kaggle.com/datasets/rhammell/ships-in-satellite-imagery.
• [3]Ashwin Acharya and Zachary Arnold.Chinese Public AI R&D Spending: Provisional Findings.Technical report, Center for Security and Emerging Technologies, December 2019.
• [4]Jiaqiu Ai, Zhilin Pei, Baidong Yao, Zhaocheng Wang, and Mengdao Xing.AIS Data Aided Rayleigh CFAR Ship Detection Algorithm of Multiple-Target Environment in SAR Images.IEEE Transactions on Aerospace and Electronic Systems, 58(2):1266–1282, April 2022.
• [5]Jiaqiu Ai, Ruitian Tian, Qiwu Luo, BoTang, and Jing Jin.Multi-Scale Rotation-Invariant Haar-Like Feature Integrated CNN-Based Ship Detection Algorithm of Multiple-Target Environment in SAR Imagery.IEEE Transactions on Geoscience and Remote Sensing, 57(12):10070–10087, December 2019.
• [6]Martina Aiello, Renata Vezzoli, and Marco Gianienetto.Object-based image analysis approach for vessel detection on optical and radar image.Journal of Applied Remote Sensing, 13(1), January 2019.
• [7]Atithee Apoorva, GopalKrishna Mishra, RashmiRanjan Sahoo, SouravKumar Bhoi, and Chittaranjan Mallick.Deep Learning-Based Ship Detection in Remote Sensing Imagery Using TensorFlow.Advances in Machine Learning and Computational Intelligence, July 2020.
• [8]Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio.Neural Machine Translation by Jointly Learning to Align and Translate.CoRR, September 2014.
• [9]Mari-Cortes Benito-Ortiz, D.Mata-Moya, M.P. Jarabo-Amores, N.del Rey-Maestre, and J.L. Barcena-Humanes.Design and Evaluation of a Artificial Intelligence Based Vessel Detection System in Pol-SAR Images.12th European Conference on Synthetic Aperture Radar, pages 1–5, 2018.
• [10]Konstantina Bereta, Raffaele Grasso, and Dimitris Zissis.Vessel Detection using Image Processing and Neural Networks.2020 IEEE International Geoscience and Remote Sensing Symposium, pages 2276–2279, 2020.
• [11]Shovakar Bhattacharjee, Palanisamy Shanmugam, and Sukhendu Das.A Deep-Learning-Based Lightweight Model for Ship Localizations in SAR Images.IEEE Access, 11:94415–94427, 2023.
• [12]Tobias Carman and Avyaya Kolhatkar.A comparison of fixed threshold CFAR and CNN ship detection methods for S-band NovaSAR images.Small Satellite Conference, August 2020.
• [13]Yang-Lang Chang, Amare Anagaw, Lena Chang, YiChun Wang, Chih-Yu Hsiao, and Wei-Hong Lee.Ship detection based on YOLOv2 for SAR imagery.Remote Sensing, 11(7):786, January 2019.
• [14]Kaiyan Chen, Ming Wu, Jiaming Liu, and Chuang Zhang.FGSD: A Dataset for Fine-Grained Ship Detection in High Resolution Satellite Images, March 2020.
• [15]Xinlei Chen and Abhinav Gupta.Webly Supervised Learning of Convolutional Networks.In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1431–1439, Santiago, Chile, December 2015. IEEE.
• [16]Zhijun Chen, Depeng Chen, Yishi Zhang, Xiaozhao Cheng, Mingyang Zhang, and Chaozhong Wu.Deep learning for autonomous ship-oriented small ship detection.Safety Science, 130:104812, October 2020.
• [17]Alina Ciocarlan and Andrei Stoian.Ship Detection in Sentinel 2 Multi-Spectral Images with Self-Supervised Learning.Remote Sensing, 13(21):4255, January 2021.
• [18]Commander, U.S. 7th Fleet Public Affairs.U.S. Navy Destroyer Conducts Freedom of Navigation Operation in the South China Sea.https://www.navy.mil/Press-Office/News-Stories/Article/3771407/us-navy-destroyer-conducts-freedom-of-navigation-operation-in-the-south-china-s/, May 2024.
• [19]Christina Corbane, Laurent Najman, Emilien Pecoul, Laurent Demagistri, and Michel Petit.A complete processing chain for ship detection using optical satellite imagery.International Journal of Remote Sensing, 31(22):5837–5854, December 2010.
• [20]Aaron WalterAvila Cordova, WilliamCondori Quispe, Remy JorgeCuba Inca, WilderNina Choquehuayta, and EvelingCastro Gutierrez.New Approaches and Tools for Ship Detection in Optical Satellite Imagery.Journal of Physics: Conference Series, 1642(1):012003, September 2020.
• [21]Rebecca Couchman-Crook, Avyaya Kolhatkar, Tobie Carman, Jess Park, James Lambert, and Claire Morris.NovaSAR and SSTL s1-4: SAR and EO data fusion.Small Satellite Conference, August 2020.
• [22]Dengxin Dai and Wen Yang.Satellite Image Classification via Two-Layer Sparse Coding With Biased Image Representation.IEEE Geoscience and Remote Sensing Letters, 8(1):173–176, January 2011.
• [23]Yuwen Deng, Donghai Guan, Yanyu Chen, Weiwei Yuan, Jiemin Ji, and Mingqiang Wei.SAR-SHIPNET: SAR-SHIP DETECTION NEURAL NETWORK VIA BIDIRECTIONAL COORDINATE ATTENTION AND MULTI-RESOLUTION FEATURE FUSION.March 2022.
• [24]Yanghua Di, Zhiguo Jiang, and Haopeng Zhang.A Public Dataset for Fine-Grained Ship Classification in Optical Remote Sensing Images.Remote Sensing, 13(4):747, January 2021.
• [25]Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, MichealYing Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang.Object detection in aerial images: A large-scale benchmark and challenges.arXiv:2102.12219 [cs], February 2021.
• [26]Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby.An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, June 2021.
• [27]Pengming Feng, Mingjie Xie, Hongning Liu, Xuanjia Zhao, Guangjun He, Xueliang Zhang, and Jian Guan.SISP: A Benchmark Dataset for Fine-grained Ship Instance Segmentation in Panchromatic Satellite Images.February 2024.
• [28]Ulrike Franke.Not smart enough: The poverty of European military thinking on artificial intelligence.\usepackage{url} \def\UrlBreaks{\do\/\do-} \usepackage{breakurl} \usepackage[breaklinks]{hyperref}, December 2019.
• [29]Antonio-Javier Gallego, Antonio Pertusa, and Pablo Gil.Automatic Ship Classification from Optical Aerial Images with Convolutional Neural Networks.Remote Sensing, 10(4):511, April 2018.
• [30]Guangshuai Gao, Qingjie Liu, and Yunhong Wang.Counting from Sky: A Large-scale Dataset for Remote Sensing Object Counting and A Benchmark Method.IEEE Transactions on Geoscience and Remote Sensing, August 2020.
• [31]Junyu Gao, Liangliang Zhao, and Xuelong Li.NWPU-MOC: A Benchmark for Fine-grained Multi-category Object Counting in Aerial Images.Journal of Latex Class Files, 14(8), August 2021.
• [32]Leonardo Gubello.Image processing with traditional and artificial intelligence techniques for ship detection.December 2022.
• [33]Ritwik Gupta, Shufan Li, Tyler Zhu, Jitendra Malik, Trevor Darrell, and Karttikeya Mangalam.xT: Nested Tokenization for Larger Context in Large Images, March 2024.
• [34]Xiyue Hou, Wei Ao, Qian Song, Jian Lai, Haipeng Wang, and Feng Xu.FUSAR-Ship: Building a high-resolution SAR-AIS matchup dataset of Gaofen-3 for ship detection and recognition.Science China Information Sciences, 63(4):140303, March 2020.
• [35]Yuxin Hu, Yini Li, and Zongxu Pan.A Dual-Polarimetric SAR Ship Detection Dataset and a Memory-Augmented Autoencoder-Based Detection Method.Sensors, 21(24):8478, January 2021.
• [36]Bogdan Iancu, Valentin Soloviev, Luca Zelioli, and Johan Lilius.ABOShips – An Inshore and Offshore Maritime Vessel Detection Dataset with Precise Annotations, February 2021.
• [37]Cosimo Ieracitano, Nadia Mammone, and FrancescoCarlo Morabito.Towards an Explainable Artificial Intelligence Approach for Ships Detection from Satellite Imagery.International Conference on Applied Intelligence and Informatics, June 2023.
• [38]Hencheng Jia and Feng Xu.Ship Detection in SAR Images with Human-in-the-Loop.
• [39]Serdar Kızılkaya, Ugur Alganci, and Elif Sertel.VHRShips: An Extensive Benchmark Dataset for Scalable Deep Learning-Based Ship Detection Applications.ISPRS International Journal of Geo-Information, 11(8):445, August 2022.
• [40]Darius Lam, Richard Kuzma, Kevin McGee, Samuel Dooley, Michael Laielli, Matthew Klaric, Yaroslav Bulatov, and Brendan McCord.xView: Objects in context in overhead imagery.arXiv:1802.07856 [cs], February 2018.
• [41]JamesM. Landreth.Through DoD’s Valley of Death: A Data-Intensive Startup’s Journey.Defense Acquisition Magazine, (Jan-Feb 2022), February 2022.
• [42]Y.Lecun, L.Bottou, Y.Bengio, and P.Haffner.Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, November 1998.
• [43]Sang-Heon Lee, Hae-Gwang Park, Ki-Hoon Kwon, Byeong-Hak Kim, MinYoung Kim, and Seung-Hyun Jeong.Accurate Ship Detection Using Electro-Optical Image-Based Satellite on Enhanced Feature and Land Awareness.Sensors, 22(23):9491, January 2022.
• [44]Songlin Lei, Dongdong Lu, Xiaolan Qiu, and Chibiao Ding.SRSDD-v1.0: A High-Resolution SAR Rotation Ship Detection Dataset.Remote Sensing, 13(24):5104, January 2021.
• [45]Boying Li, Bin Liu, Lanqing Huang, Weiwei Guo, Zenghui Zhang, and Wenxian Yu.OpenSARShip 2.0: A large-volume dataset for deeper interpretation of ship targets in Sentinel-1 imagery.In 2017 SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA), pages 1–5, November 2017.
• [46]Jianwei Li, Changwen Qu, and Jiaqi Shao.Ship detection in SAR images based on an improved faster R-CNN.In 2017 SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA), pages 1–6, November 2017.
• [47]KeLi, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han.Object detection in optical remote sensing images: A survey and a new benchmark.ISPRS Journal of Photogrammetry and Remote Sensing, 159:296–307, January 2020.
• [48]Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C.Lawrence Zitnick.Microsoft COCO: Common Objects in Context.In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing.
• [49]Zikun Liu, Jingao Hu, Lubin Weng, and Yiping Yang.Rotated region based CNN for ship detection.In 2017 IEEE International Conference on Image Processing (ICIP), pages 900–904, September 2017.
• [50]Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang.A High Resolution Optical Satellite Image Dataset for Ship Recognition and Some New Baselines:.In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods, pages 324–331, Porto, Portugal, 2017. SCITEPRESS - Science and Technology Publications.
• [51]Xiaolin Ma, Junkai Cheng, Aihua Li, Yuhua Zhang, and Zhilong Lin.AMANet: Advancing SAR Ship Detection with Adaptive Multi-Hierarchical Attention Network.IEEE Access, 11, 2023.
• [52]Armando Marino, MariaJ. Sanjuan-Ferrer, Irena Hajnsek, and Kazuo Ouchi.Ship detection with spectral analysis of synthetic aperture radar: A comparison of new and well-known algorithms.Remote Sensing, 7(5):5416–5439, May 2015.
• [53]Aristides Milios, Konstantina Bereta, Konstantinos Chatzikokolakis, Dimitris Zissis, and Stan Matwin.Automatic fusion of satellite imagery and AIS data for vessel detection.In 2019 22th International Conference on Information Fusion (FUSION), pages 1–5, July 2019.
• [54]NarendraKumar Mishra, Ashok Kumar, and Kishor Choudhury.Deep Convolutional Neural Network based Ship Images Classification.Defence Science Journal, 71(2):200–208, March 2021.
• [55]Fernando Paolo, Tsu-tingTim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon.xView3-SAR: Detecting Dark Fishing Activity Using Synthetic Aperture Radar Imagery.NeurIPS Proceedings, 2022.
• [56]Krishna Patel, Chintan Bhatt, and PierLuigi Mazzeo.Deep Learning-Based Automatic Detection of Ships: An Experimental Study Using Satellite Images.J. Imaging, 8(7), June 2022.
• [57]Ramona Pelich, Marco Chini, Renaud Hostache, Patrick Matgen, Carlos Lopez-Martinez, Miguel Nuevo, Philippe Ries, and Gerd Eiden.Large-scale automatic vessel monitoring based on dual-polarization sentinel-1 and AIS data.Remote Sensing, 11(9):1078, January 2019.
• [58]Xiaoman Qi, Panpan Zhu, Yuebin Wang, Liqiang Zhang, Junhuan Peng, Mengfan Wu, Jialong Chen, Xudong Zhao, Ning Zang, and P.Takis Mathiopoulos.MLRSNet: A multi-label high spatial resolution remote sensing dataset for semantic scene understanding.ISPRS Journal of Photogrammetry and Remote Sensing, 169:337–350, November 2020.
• [59]Stuart Russell and Peter Norvig.Artificial Intelligence: A Modern Approach.Pearson, 4 edition, 2021.
• [60]AbraarRaza Samar and Adeel Mumtaz.Enhancing Performance of IR Ship Detection with Baseline AI Models over a new Benchmark.IEEE Asia-Pacific Conference on Computer Science and Data Engineering, pages 1–6, 2021.
• [61]Shreya Shankar, Rolando Garcia, JosephM. Hellerstein, and AdityaG. Parameswaran.Operationalizing Machine Learning: An Interview Study.https://arxiv.org/abs/2209.09125v1, September 2022.
• [62]Zhenfeng Shao, Wenjing Wu, Zhongyuan Wang, Wan Du, and Chengyuan Li.SeaShips: A large-scale precisely annotated dataset for ship detection.IEEE Transactions on Multimedia, 20(10):2593–2604, October 2018.
• [63]Tom Shugart.Deterring the Powerful Enemy.https://www.cnas.org/publications/congressional-testimony/deterring-the-powerful-enemy, March 2024.
• [64]Jiexiong Tang, Chenwei Deng, Guang-Bin Huang, and Baojun Zhao.Compressed-domain ship detection on spaceborne optical image using deep neural network and extreme learning machine.IEEE Transactions on Geoscience and Remote Sensing, 53(3):1174–1185, March 2015.
• [65]HildeM. Toonen and SimonR. Bush.The digital frontiers of fisheries governance: Fish attraction devices, drones and satellites.Journal of Environmental Policy & Planning, 22(1):125–137, January 2020.
• [66]U.S. 7th Fleet Public Affairs.U.S., Japan Conduct Bilateral Exercise in the South China Sea.https://www.cpf.navy.mil/Newsroom/News/Article/3679868/us-japan-conduct-bilateral-exercise-in-the-south-china-sea/, February 2024.
• [67]Chao Wang, Shaofeng Jiang, Hong Zhang, Fan Wu, and BoZhang.Ship Detection for High-Resolution SAR Images Based on Feature Analysis.IEEE Geoscience and Remote Sensing Letters, 11(1):119–123, January 2014.
• [68]Chenwei Wang, Jifang Pei, Siyi Luo, Weibo Huo, Yulin Huang, Yin Zhang, and Jianyu Yang.SAR Ship Target Recognition Via Multi-Scale Feature Attention and Adaptive-Weighed Classifier.August 2023.
• [69]Yuanyuan Wang, Chao Wang, Hong Zhang, Yingbo Dong, and Sisi Wei.A SAR dataset of ship detection for deep learning under complex backgrounds.Remote Sensing, 11(7):765, January 2019.
• [70]WANG Zhirui, KANG Yuzhuo, ZENG Xuan, WANG Yuelei, ZHANG Ting, and SUN Xian.SAR-AIRcraft-1.0: High-resolution SAR Aircraft Detection and Recognition Dataset.Journal of Radars, 12(4), August 2023.
• [71]Shunjun Wei, Xiangfeng Zeng, Qizhe Qu, Mou Wang, Hao Su, and Jun Shi.HRSID: A High-Resolution SAR Images Dataset for Ship Detection and Instance Segmentation.IEEE Access, 8:120234–120254, 2020.
• [72]BrianA. Wong, Christian Thomas, and Patrick Halpin.Automating offshore infrastructure extractions using synthetic aperture radar & google earth engine.Remote Sensing of Environment, 233:111412, November 2019.
• [73]Jixiang Wu, Zongxu Pan, Bin Lei, and Yuxin Hu.LR-TSDet: Towards Tiny Ship Detection in Low-Resolution Remote Sensing Images.Remote Sensing, 13(19):3890, January 2021.
• [74]Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang.DOTA: A large-scale dataset for object detection in aerial images.In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974–3983, Salt Lake City, UT, June 2018. IEEE.
• [75]Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu.AID: A Benchmark Data Set for Performance Evaluation of Aerial Scene Classification.IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965–3981, July 2017.
• [76]S.U.N. Xian, Wang Zhirui, S.U.N. Yuanrui, Diao Wenhui, Zhang Yue, and F.U. Kun.AIR-SARShip-1.0: High-resolution SAR Ship Detection Dataset (in English).Journal of Radars, 8(6):852–863, December 2019.
• [77]Hang Xiao, Xin Wang, and Peng Zhao.Satellite Image Recognition for Smart Ships Using A Convolutional Neural Networks Algorithm.International Journal of Decision Science, 10(2):85–91, June 2019.
• [78]Guang Yang, BoLi, Shufan Ji, Feng Gao, and Qizhi Xu.Ship detection from optical satellite images based on sea surface analysis.IEEE Geoscience and Remote Sensing Letters, 11(3):641–645, March 2014.
• [79]Guangmiao Zeng, Wanneng Yu, Rongjie Wang, and Anhui Lin.Research on Mosaic Image Data Enhancement for Overlapping Ship Targets.May 2021.
• [80]Chuan Zhang, Gui Gao, Linlin Zhang, C.Chen, S.Gao, Libo Yao, Qilin Bai, and Shiquan Gao.A Novel Full-Polarization SAR Images Ship Detector Based on the Scattering Mechanisms and the Wave Polarization Anisotropy.ISPRS Journal of Photogrammetry and Remote Sensing, December 2021.
• [81]Feng Zhang, Xueying Wang, Shilin Zhou, Yingqian Wang, and YiHou.Arbitrary-Oriented Ship Detection through Center-Head Point Extraction.IEEE Transactions on Geoscience and Remote Sensing, October 2021.
• [82]Tianwen Zhang, Xiaoling Zhang, Xiao Ke, XuZhan, Jun Shi, Shunjun Wei, Dece Pan, Jianwei Li, Hao Su, Yue Zhou, and Durga Kumar.LS-SSDD-v1.0: A deep learning dataset dedicated to small ship detection from large-scale sentinel-1 SAR images.Remote Sensing, 12(18):2997, January 2020.
• [83]Tianwen Zhang, Xiaoling Zhang, Jianwei Li, Xiaowo Xu, Baoyou Wang, XuZhan, Yanqin Xu, Xiao Ke, Tianjiao Zeng, Hao Su, Israr Ahmad, Dece Pan, Chang Liu, Yue Zhou, Jun Shi, and Shunjun Wei.SAR Ship Detection Dataset (SSDD): Official Release and Comprehensive Data Analysis.Remote Sensing, 13(18):3690, January 2021.
• [84]Yuanlin Zhang, Yuan Yuan, Yachuang Feng, and Xiaoqiang Lu.Hierarchical and robust convolutional neural network for very high-resolution remote sensing object detection.IEEE Transactions on Geoscience and Remote Sensing, 57(8):5535–5548, August 2019.
• [85]Zenghui Zhang, Weiwei Guo, Shengnan Zhu, and Wenxian Yu.Toward arbitrary-oriented ship detection with rotated region proposal and discrimination networks.IEEE Geoscience and Remote Sensing Letters, 15(11):1745–1749, November 2018.
• [86]Zhengning Zhang, Lin Zhang, Yue Wang, Pengming Feng, and Ran He.ShipRSImageNet: A Large-Scale Fine-Grained Dataset for Ship Detection in High-Resolution Optical Remote Sensing Images.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:8458–8472, 2021.
• [87]Dandan Zhao, Zhe Zhang, Dongdong Lu, Jian Kang, Xiaolan Qiu, and Yirong Wu.CVGG-Net: Ship Recognition for SAR Images Based on Complex-Valued Convolutional Neural Network.
• [88]Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl.Objects as Points.arXiv:1904.07850 [cs], April 2019.
• [89]Changren Zhu, Hui Zhou, Runsheng Wang, and Jun Guo.A novel hierarchical method of ship detection from spaceborne optical image based on shape and texture features.IEEE Transactions on Geoscience and Remote Sensing, 48(9):3446–3456, September 2010.