🌲 ForestSentry

On-device forest monitoring for the rangers who actually walk the trails.

Features · Getting Started · Usage · Report Bug · Request Feature

---

ForestSentry is an offline-first React Native app for forest rangers and conservation researchers to tag trees, classify leaf health on-device with TensorFlow Lite, log illegal-logging incidents, and monitor protected zones with PostGIS. Built with WWF Pakistan for remote forests like Nathia Gali — where signal disappears but the case file still has to hold up. Open-source, Expo SDK 54, TypeScript strict, MIT.

🚧 Active development — V1 scaffold is feature-complete (tree registry, on-device classifier, incident reporting, zone polygons, offline sync, active-learning loop). Field data collection and production model training are in progress. See the Roadmap.

---

✨ Features

	Feature	Description
📍	20-second tree registry	GPS auto-captures; species, girth, and height in a single form. Photo optional.
🍃	On-device leaf-health AI	Bundled PlantVillage MobileNet (39-class → 4-bucket verdict) via react-native-fast-tflite. Deterministic HSV fallback keeps the flow working without the native module.
🔁	Active-learning loop	Operators confirm or correct every verdict; corrections feed a Supabase pipeline that retrains MobileNetV2 and ships improved models OTA — no app-store release.
🚨	Incident reporting	Observation / minor / serious / critical severity, with photo + GPS + notes. Queues offline.
🗺️	PostGIS zone monitoring	Draw polygons on the map; real-time inside/outside detection for protected zones.
📈	Trend projection	Linear regression + EMA over inspection history, rendered with Skia.
📤	Dataset export	CSV for spreadsheets, GeoJSON for QGIS — one tap.
🧭	Offline-first by design	SQLite is the source of truth; the sync engine drains to Supabase the moment signal returns.
♿	Field-grade UX	High-visibility sun mode, system/light/dark themes, skeleton loading, full a11y (role + label + hint, reduced-motion aware).

---

🛠️ Tech Stack

Category	Technology
Framework	Expo SDK 54 (New Architecture), React Native 0.81
Language	TypeScript (strict + noUncheckedIndexedAccess)
Navigation	expo-router (file-based, typed routes)
UI / Motion	react-native-reanimated 4, @shopify/flash-list v2, @shopify/react-native-skia, expo-image
AI inference	react-native-fast-tflite + jpeg-js (JPEG → RGB tensor)
Training	TensorFlow 2.16 / Keras, MobileNetV2 transfer learning, int8 TFLite (off-device, Colab)
Local data	expo-sqlite (WAL), @tanstack/react-query + AsyncStorage persister
State / Forms	zustand, react-hook-form + zod
Maps	react-native-maps (native MapView + polygons)
Backend	Supabase — Postgres + PostGIS + Storage, RLS-enforced
Auth	@supabase/supabase-js email OTP (6/8-digit code, no deep links)

---

🚀 Getting Started

Prerequisites

• Node.js >= 20 (22 recommended)
• Git
• Xcode 16 + iOS Simulator, or Android Studio + a Pixel emulator
• A Supabase project (free tier is fine)

Installation

git clone https://github.com/aashir-athar/forest-sentry.git
cd forest-sentry/mobile
npm install

Configure

cp .env.example .env
# fill in EXPO_PUBLIC_SUPABASE_URL and EXPO_PUBLIC_SUPABASE_ANON_KEY

Run

npx expo start

For full Supabase setup, EAS builds, and store submission, see zero-to-deploy.md.

---

📖 Usage

All scripts run inside mobile/.

npm run start      # Boot the Expo dev server
npm run ios        # Open the iOS simulator
npm run android    # Open the Android emulator
npm run web        # Web preview (for inspection)
npm run lint       # Run the Expo ESLint config
npx tsc --noEmit   # Strict TypeScript pass

Active-learning loop: classify, correct, retrain, ship OTA

1. Settings → Training contribution → ON (opt-in, off by default).
2. Tag a tree and open Inspect a leaf → take a photo.
3. The on-device classifier predicts a verdict (healthy / stressed / diseased / pest-damaged).
4. Confirm it, or tap a different pill to correct it — the diff is the training signal.
5. Local SQLite records predicted_label, corrected_label, and model_version; the sync engine pushes it to public.training_samples_leaf when signal returns.
6. Retrain MobileNetV2 with backend/training/train_leaf_health.py, then promote the new model:

select public.wwf_model_promote('leaf-health-v1.1.0');

Every device downloads the new .tflite on next launch via expo-file-system. The bundled placeholder remains the offline-first cold-install fallback.

Swap in your own WWF-trained classifier

Overwrite mobile/assets/models/leaf-health.tflite with your .tflite, then update the modelClasses array and BUCKETS table in mobile/src/features/inspections/labels.ts to match your class order and bucket assignments. The classifier reads the model's declared input shape and dtype at load time — no other code changes required.

---

🗺️ Roadmap

• V1 — Tree registry, on-device classifier, incident reporting, zone polygons, offline sync
• V1 — Supabase backend with RLS, PostGIS, role helpers
• V1 — CSV / GeoJSON export, trend projection, high-visibility sun mode
• V1 — Active-learning loop: correction UI, training opt-in, model_versions registry, OTA downloads, Colab training script
• V1.1 — Logging-detection model trainer (once a field corpus exists)
• V1.1 — Vision Camera frame-processor TFLite (live leaf scan, no shutter)
• V1.2 — Multi-language UI (Urdu, Pashto)
• V2 — Predictive tree-health time series + incident-hotspot anomaly models
• V2 — WatermelonDB and cluster rendering for very large datasets

---

🤝 Contributing

Contributions are very welcome — there's a good first issue label for newcomers, and the codebase is intentionally readable.

1. Fork the repo
2. Create a branch (git checkout -b feat/your-thing)
3. Commit using Conventional Commits (feat:, fix:, docs:)
4. Push and open a PR against main

See CONTRIBUTING.md for details. If you're a botanist or forest scientist, the highest-impact contribution is a real .tflite leaf-health model — see mobile/assets/models/README.md.

---

📄 License

Distributed under the MIT License. See LICENSE for details.

---

👤 Author

Aashir Athar

---

If ForestSentry helps your conservation work, consider leaving a ⭐ — it helps other rangers and researchers find it.

_{Keywords: react native offline-first app · expo sdk 54 starter · tensorflow lite leaf classifier mobile · on-device plant disease classifier · plantvillage mobilenet react native · active learning loop react native · supabase postgis react native · conservation field app open source · WWF forest monitoring tool · expo-router file-based navigation · react-native-fast-tflite example · on-device ML conservation · GPS tree registry · illegal logging incident reporting}

_{Built by aashir-athar · MIT Licensed}