revised_roadmap_v2.md

A5 Revised Roadmap v2: AgMultida Alignment

Critical Misalignment Fix

Current repo state: 'water stress binary'. Target state: 3 rice targets (Blast, BPH, N-Def).

Architecture & Data

• Stack: Keep EfficientNetB3 + GRU + CrossAttn.
• Model Head: Switch from binary to multi-task head (3 classes).
• Labeling Strategy: Multi-label expert-aligned.
• Green Ag Goals: Align metrics with Water + IPM (Integrated Pest Management) optimization.

Fast-Track 14-Day Timeline

Phase 0: Triage & Config (Days 1-2)

• Purge binary classification logic.
• Update config files for 3-class target.

Phase 1: Data Strategy (Days 3-5)

• Implement multi-label expert-aligned pipelines.
• Ingest Blast, BPH, and N-Def datasets.

Phase 2: Model Adaptation (Days 6-8)

• Retain EfficientNetB3+GRU+CrossAttn backbone.
• Attach and initialize 3-class multi-task head.

Phase 3: Green Ag Integration (Days 9-10)

• Embed Water and IPM constraint metrics into loss/evaluation.
• Tune for resource efficiency.

Phase 4: Training & Validation (Days 11-12)

• Execute training loop on 3-target data.
• Validate multi-class performance metrics.

Phase 5: Handoff & Freeze (Days 13-14)

• Finalize model weights.
• Verify 14-day completion gates.