Lt.j.g. Mustafa Emre Aydınlık wasborn in İzmir, Türkiye. He earned his commission and a BS in Electrical & Electronics Engineering from the Turkish Naval Academy in 2021. He then served two years as a Branch Officer on Turkish Warships.
Aydınlık arrived at the Naval Postgraduate School in 2023. During his master’s program he completed the Space Systems Fundamentals, Signal Processing, and Senior Electronic Warfare Engineering certificate series while maintaining distinction level grades.
His thesis, “Enhancing SAR Target Recognition with GAN Based Data Augmentation”, bridges radar engineering and artificial intelligence. By training a Generative Adversarial Network on the MSTAR benchmark and integrating the synthetic imagery as a dedicated “ninth class,” he boosted convolutional network validation accuracy from 94 % to≈ 99 %, delivering a low-cost blueprint for all weather Automatic Target Recognition when real data are scarce. Aydınlık’s work directly supports NATO interoperability initiatives and has been briefed to Turkish Navy and U.S. Navy.
The work was motivated by a clear operational gap: modern Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) algorithms (e.g., Convolutional Neural Network, or CNNs) struggle when training data is scarce. As noted in the thesis, “deep networks generally need large datasets but for SAR imagery obtaining data is difficult. For instance, MSTAR Dataset has only a few hundred images per target class which may cause CNNs to overfit.” In practice, collecting real SAR images of every military vehicle(especially adversary models) is expensive or impossible. In fact, the MSTAR benchmark itself contains vehicles “belonging to non-allied countries,” underscoring the data gaps in allied forces’ training sets.
The thesis asks whether Generative Adversarial Networks (GANs) can synthetically augment limited SAR datasets and thereby improve classification. By generating realistic synthetic SAR images, the work aims to increase sample diversity without additional flights or sensors. In short, the specific problem was SAR data scarcity for CNN training, and the thesis explores GAN-based augmentation to bridge that gap.
Improving radar ATR supports interoperability by strengthening shared situational awareness across allies. NATO and partner nations are actively investing in multinational SAR intelligence sharing. For example, NATO’s recent Alliance Persistent Surveillance from Space(APSS) initiative brings commercial SAR data (from providers like ICEYE) into the NATO intelligence enterprise. This kind of alliance-wide surveillance depends on reliable ATR: if each country’s algorithms can robustly classify objects in SAR imagery, then the coalition gains a common operating picture.Similarly, NATO’s MAJIIC program explicitly targets coalition ISR interoperability. MAJIIC’s Concept of Employment outlines network-enabled integration of SAR, GMTI, and EO imagery for NATO operations. By demonstrating that CNN classifiers can be made more accurate and general (viaGAN-augmented training), this research directly supports those NATO goals:better ATR means that Turkish, US, and other systems can exchange SAR threat assessments on equal footing.
A. Shared SAR Intelligence: A secure coalition requires shared space/air ISR. Recent announcements (e.g. ICEYE supplying SAR data toNATO) highlight alliance efforts to “increase space-based intelligence sharing across the Alliance.” Robust target classifiers make that shared data actionable for all partners.
B. Coalition ISTAR Integration: NATO doctrine emphasizes integrating allied sensor streams. The MAJIIC project explicitly includes SAR imagery as a key ISR product for multinational missions. This thesis’ techniques enhanceSAR image classification, complementing such coalition ISTAR architectures.
C. Standards and Interoperability: Improved SAR recognition algorithms (grounded in open research like this thesis) help ensure thatTurkish and U.S. systems speak the same language when classifying threats.
D. Common Threat Awareness: Crucially, the method was validated on the MSTAR dataset, which includes threat vehicles (some from potential adversaries). A more general classifier means allied forces can more reliably identify those targets in shared operations or exercises.
The Naval Postgraduate School provided the computational tools and mentorship essential for this work. NPS’ faculty and research centers are deeply engaged in AI for defense. For instance,NPS’ Center for Multi-INT Studies has applied convolutional neural networks and other machine-learning models to imagery and signals for intelligence tasks and previous NPS theses have explored GANs for classification. This environment, access to high-performance computing (GPU clusters, MATLAB toolboxes, etc.) and expert advisors in signal processing and machine learning, enabled rapid prototyping of GAN and CNN workflows. It also provided access to key datasets (like MSTAR)and domain expertise in radar. In short, NPS’ facilities and faculty expertise in both radar signal processing and deep learning underpinned the project’s success.
Looking forward, future students and faculty can build on this foundation in several ways:
A. Class-Conditional GANs: The thesis recommends using class-conditional GANs (CGANs) to generate images specific to each target class. This would eliminate uncertainty about which synthetic images correspond to which real class, potentially making augmentation more effective. A CGAN could produce, say, synthetic tanks versus trucks with explicit labels, enhancing training integrity.
B. Improved GAN Realism: Work can improve GAN architectures for SAR physics. For example, adding a high-frequency-focused discriminator can yield synthetic images with more realistic speckle noise. Future research should incorporate SAR-specific constraints so that generated images better mimic real radar distortions, further boosting classifier generalization.
C. Maintaining Label Integrity: Our key finding was that synthetic data should be carefully labeled. The thesis found that treating all GAN-generated images as a separate “fake” class dramatically improved result.Future experiments should explore hybrid strategies (e.g. mix real and synthetic samples of each class with confidence scores) but must preserve label integrity to avoid confusing the CNN.
D. Expanded Datasets and Domains: Building on the GAN approach, future work could fuse data from additional sensors (multistatic SAR, different frequencies, or polarimetric channels) and international datasources. As more coalition datasets become available, combining real and generated data across countries would further enhance interoperability.
