PesteVision: Offline pest spotter for rice and vegetable plots

Problem Statement

Smallholders lose 15-30% of yield to misidentified pests: spray the wrong chemical, spray too late, or ignore a deficiency as “normal.” The municipal agriculturist covers dozens of barangays and can’t visit every plot. Facebook advice is generic and sometimes pushes over-spraying. Signal in the field is often zero.

Proposed Solution

Point, shoot, get an answer: the app classifies common rice/vegetable pests and nutrient issues from a leaf photo, gives a Tagalog explanation with confidence (“Mukhang rice blast, 78% — tingnan kung may diamond spots”), and lists low-cost next steps plus when to escalate to the MAO. Field log stays on-device; farmers can show a week of photos at the agri office without needing cloud.

Target Market

• Primary users: Rice and gulay smallholders, 4H youth farmers.
• Deployers: MAO technicians, coops, agri supply shops that want honest advice.

How AI Is Used

• On-device vision: Quantized classifier for ~25 local pest/deficiency classes; works with no signal.
• Explainer LM: Turns label + confidence into plain Tagalog steps; avoids pesticide brand endorsement.
• Why AI: Visual identification needs vision; text rules can’t cover leaf variation.

Independent Deployment Angle

• No cloud, no carrier, no venue, no headset. Runs on mid-range phones fully offline.
• Model and photo packs update via SD card at the MAO office.
• Any agri student can add new pest examples with 50 photos.

Demo Scenario (3 minutes for judges)

1. Airplane mode. Photograph a printed diseased-leaf sample.
2. App returns top-3 guesses with confidence and Tagalog notes.
3. Show a healthy leaf for contrast — honest “mukhang healthy” verdict.
4. Open the field log: 5 dated entries ready to show the MAO.
5. Close: “Sa bukid mismo, kahit walang signal.”

30-Hour Build Plan

Suggested stack: Flutter + TFLite MobileNet/YOLO-tiny, on-device Tagalog explainer TinyLM, SQLite field log. Public data / APIs: Open pest image sets, DA/PhilRice pamphlets (public guidance).

Risks & Privacy Notes

• Wrong spray advice: Never names brands or doses; always shows confidence and “ask MAO if below 70%.” Demo uses printed samples, no real farm geo.
• Privacy: Photos stay on-device; no farm location uploaded. Complies with data-minimization.
• IP: Only open/CC datasets plus team photos; licenses listed.

AI Use Disclosure (draft)

Uses a quantized vision classifier plus a small on-device explainer LM. AI coding assistants helped with training scripts. All demo verdicts were checked against MAO pamphlets by the team.

Rubric Scoring