Precision Agriculture DLI 2026 Best Poster · 200 FPS ONNX Published Suite + Manuscript Under Review

CropGuardNet & DGS-Bench: Leak-Free Field Crop Pathology & Auditing

Leak-free UUID and 64-bit pHash deduplication across 35,420 unique leaf images, sub-megabyte CropGuardNet ONNX inference (481.6K params, 5.00 ms), DGS-Bench field gap auditing, and DSA 2026 data-efficiency bounds.

CropGuardNet quad-method explainable AI attribution across eight crop species

Scientific Context & Approach

Prior plant disease detectors report inflated >99% accuracies driven by augmented duplicate leakage across splits and break catastrophically when evaluated on authentic smallholder field captures.

This core research thrust pairs leak-free multi-crop modeling (CropGuardNet across 8 crops and 38 classes) with formal deployment auditing (DGS-Bench Best Poster at DLI 2026), few-shot budget scaling and DCGAN failure proofs at DSA 2026 Kampala, and offline TFLite INT8 Android tooling.

Core Technical Objectives

  • Eliminate offline-augmented train/test contamination via specimen UUID grouping and 64-bit DCT perceptual hashing.
  • Engineer sub-0.5M parameter hybrid CNN-Transformer architectures running at 200 FPS in single-thread CPU ONNX.
  • Quantify field deployment collapse across minority diseases and publish reproducible Kaggle T4 benchmark harnesses.

Verified Deliverables & Outputs

  • CropGuardNet framework & 80,238-duplicate purge audit
  • DGS-Bench Best Poster deployment gap benchmark (DLI 2026 Lagos)
  • Three DSA 2026 Kampala papers + DeepLabV3+/MobileViT offline Maize field app