In September 2026 I posted on LinkedIn about the role of photographs in Digital Green’s FarmerChat system, the scale of India’s digital agriculture infrastructure, and broader questions about AI reliability, data quality, and the need for human judgment.
- Total photos analyzed: 1.16 million across four countries.
Key themes that emerged were an image‑quality bottleneck, a three‑stage decision pipeline, data‑quality challenges, and the necessity of human‑in‑the‑loop review.
- Image‑quality bottleneck: Nearly half of submitted images in India are rejected before any diagnostic AI can run.
I decided to adopt an image‑quality gate, implement the three‑stage pipeline, prioritize representative field datasets, and integrate human review because these steps address upstream waste, improve downstream accuracy, and preserve trusted farmer interactions.
- Adopt image‑quality gate to reduce wasted model cycles.
Open questions I posed included drivers of the high rejection rate in India versus Kenya, how to integrate multimodal sensing, optimal usability thresholds, scaling human review, and governance for maintaining representative datasets.
- Implement three‑stage pipeline to allow targeted feedback to farmers.
- Prioritize representative, multimodal field datasets for robust models.
- Integrate human review for low‑confidence cases and to provide trusted recommendations.
- Three‑stage decision system: (a) image usability classifier, (b) crop classification, (c) disease/pest detection, enabling targeted recapture requests.
- Data quality challenges: fragmented datasets, inconsistent annotations, inadequate metadata, limited validation across locations/seasons.
- Human‑in‑the‑loop necessity: judgment, contextual understanding, and communication back to farmers remain critical.
- India submissions: 389,597 photos; 45 % rejected before diagnosis.
- Kenya rejection rate: 18.3 %.
- Digital agriculture scale: >10.31 crore farmer IDs, 31.3 crore surveyed plots, >10,000 extension workers, 73 crops, 436 pests, >72 lakh farmer queries answered by #BharatVistaar.
- Krishi Vigyan Kendra network: 731 KVKs available as frontline institutions.