Research
SAR Automatic Target Recognition — Generalization Study
2025–2026 · Under Review, IEEE Geoscience and Remote Sensing Letters
Systematic study of SAR ATR generalization across depression angles, sensor configurations, and datasets using the ATRNet-STAR dataset (40 vehicle classes, 4 depression angles) and the MSTAR benchmark (10 military vehicle classes).
Six contributions:
- Depression angle ablation: adding a single 60° view to a 15°-only baseline improves EOC accuracy by +30.4 percentage points
- Data efficiency: multi-angle training with as few as 25 images per vehicle per angle surpasses models trained on thousands of single-angle images
- Neither angle-weighted sampling nor test-time augmentation improves EOC performance — TTA reduces accuracy by 24.3 pp — confirming angle invariance requires real multi-angle data
- Few-shot prototypical classifier achieving 98%+ accuracy with only 5 reference images per class
- Cross-dataset transfer revealing a 38 pp SAR-to-SAR domain gap
- DINOv2 optical pretrained features achieve only 6.1% on SAR ATR — directly motivating domain-specific SAR backbone development
Preprint: https://doi.org/10.5281/zenodo.20492213
SAR Ship Type Classification from Sentinel-1 SLC Imagery
2025–2026 · Under Review, IEEE
Ship type classification from SAR imagery addressing two underappreciated problems: noisy AIS vessel type codes used as ground truth, and predominant reliance on GRD products when free SLC data with finer resolution is available from the same sensor.
Key contributions:
- Multi-signal label resolver combining AIS type code, ship length, Global Fishing Watch classification, and vessel name keywords into a confidence-weighted vote — resolving 87.3% of 2,088 vessels to reliable type labels without manual annotation
- DINOv2 ViT-S/14 classifier trained on resolved labels comparing SLC and GRD chips under identical experimental conditions
- Public release of pipeline, labels, and 4,306 dual-polarisation SLC chips covering 2,088 unique vessels in the Gulf of Mexico
Preprint: https://doi.org/10.5281/zenodo.20492437
ShayaSAR — General SAR Foundation Model (In Progress)
2026 · Target: arXiv + Hugging Face, August 2026
DINOv2 fine-tuned on the Sentinel-1 SAR channel of SSL4EO-S12 v1.1 — 246,144 globally distributed locations, four seasonal timestamps, CC-BY-4.0 license. Filling a documented gap: no public SAR-only DINOv2 backbone with clean civilian provenance currently exists. Downstream evaluation on MSTAR, BigEarthNet-SAR, and change detection benchmarks.
George B. Moody PhysioNet Challenge 2026 (Active)
2026 · Screening for Cognitive Impairment During Sleep Studies
Developing open-source algorithms using polysomnography (PSG) recordings — including EEG, ECG, and other physiological signals — to predict future diagnoses of cognitive impairment. Training data from five US institutions: Beth Israel Deaconess Medical Center, Emory University, Kaiser Permanente, Massachusetts General Brigham, and Stanford University.
Directly extends HP/Agilent ECG signal processing background to modern multi-signal biosignal ML in a clinical research context.
Challenge site: https://moody-challenge.physionet.org/2026/
MIMIC-IV Cardiac Readmission Prediction (In Progress)
2026 · Credentialing in Progress
Planned: 30-day readmission prediction for heart failure patients using MIMIC-IV clinical data. XGBoost classifier with SHAP feature-level explainability for clinical risk stratification. Targeting medRxiv preprint and GitHub reproducible release. Relevant to CMS/HHS priorities and MITRE Health FFRDC work.
Lichen Biodiversity Informatics & ML Classification
Nov 2024 – Present · Arizona State University Research Affiliate
Collaborating with Dr. Frank Bungartz, Collections Manager of Lichens at Arizona State University, on a multi-component research program spanning database engineering, phylogenetics, and deep learning image classification.
ML Image Classification Pipeline
BioCLIP ViT-L/14 + ArcFace metric learning classifier for species-level lichen identification from images. Two-stage training: large-scale noisy pretraining on iNaturalist aggregated imagery, followed by fine-tuning on expert-labeled Lichen Consortium data (~15,000 images across ~10,000 species). Approximately 95% top-1 accuracy. Pending publication with Dr. Frank Bungartz.
- UMAP + HDBSCAN clustering pipeline over LIAS DELTA morphological data (10,709 species, 880 chemical compounds)
- Interactive three-panel browser-based cluster viewer for exploratory analysis
- Clustering analysis revealed clean photobiont-type separation
- Embedding ensemble: BioCLIP / DINOv2 / CLIP
Phylogenetic Analysis Pipeline
- Parsed ~51K GenBank DNA records (8 loci: ITS, 18S, 28S, RPB2, and others)
- MAFFT alignment, trimAl trimming, IQ-TREE2 maximum likelihood inference
- Mapping LIAS morphological and chemical trait data onto inferred phylogenetic trees
Database Engineering & Taxonomy Cleanup
PostgreSQL database (fungix schema) integrating LIAS DELTA, Mycobank, and Index Fungorum:
- Removed 35,000 non-lichen taxa from the Consortium database
- Added 60,000 new validated taxa from Mycobank and Index Fungorum
- Updated all 200,000 taxa records with external source references
- Contributor to Symbiota open-source platform
PCG Heart Sound Classification (Investigating)
Investigation of phonocardiogram classification using mel-spectrogram + Vision Transformer architectures and pretrained audio foundation models (HuBERT, Wav2Vec 2.0) — extending ECG signal processing expertise to acoustic cardiac signals.
Affiliations
- ASU Research Affiliate — Arizona State University
- Member — AFCEA Lexington-Concord Chapter
- Registered Expert — GLG Expert Network
- Registered Consultant — Catalant
- SAM.gov Registered Vendor — UEI YK29WQZ22EM3
- Participant — George B. Moody PhysioNet Challenge 2026
- Member — American Bryological and Lichenological Society (ABLS)
- Contributor — Consortium of North American Lichen Herbaria (CNALH) / Symbiota
- Registered — ODSC AI East 2026, Boston MA