Agrisynthia

Agricultural analysis platform for orchards. Counts fruit on trees from field and drone imagery using YOLOv7, tracks field health from drone orthophotos and Sentinel-2 satellite passes, estimates yield, and generates PDF and Excel reports.

The product interface is Turkish first. Source strings are Turkish and English is a translation, so a fresh install renders in Turkish without any translation catalog compiled.

What it does

Fruit detection. Upload a photo of a tree, get a count. Six models: mandarin (mandalina), apple (elma), pear (armut), peach (seftale), pomegranate (nar), and tree counting (agac). Runs synchronously or as a background job with live progress over Server-Sent Events. Results cache in Redis keyed by image hash, so re-uploading the same photo is free.

Drone mapping. Upload orthophotos, or let NodeODM build one from raw flight images. Roughly 25 vegetation indices (NDVI, GLI, VARI, SAVI, EVI, NDRE and more) render as map layers. Stress zones and tree density come out as GeoJSON.

Satellite monitoring. Sentinel-2 NDVI pulled through the Earth Search STAC API, one reading per field per scene date, refreshed weekly. Email alerts fire when a field drops below a configured NDVI threshold.

Yield and recommendations. Detection counts, tree age, and NDVI feed a yield estimate. Stress zones turn into ranked agronomic actions.

Reports. PDF and Excel for detections and drone analyses, on demand or on a daily, weekly, or monthly email schedule.

Accounts. Registration, email verification, TOTP two-factor with backup codes, password reset, profile and notification settings.

Public site. Turkish marketing site with blog, pricing, KVKK and privacy pages, a newsletter signup, and a support chatbot backed by the Anthropic API. English is served under /en/.

Requirements

• Docker 24+ and Docker Compose v2, or Python 3.10 for a local install
• NVIDIA GPU with CUDA 11.8+ for practical inference speed. The stack starts without one and falls back to CPU, which is very slow.
• NVIDIA Container Toolkit on the host, for the Docker path
• 8 GB+ GPU VRAM recommended
• About 3 GB of disk for model weights

Quick start with Docker

git clone https://github.com/ChronoCoders/agrisynthia.git
cd agrisynthia
cp deploy/env.production.template .env    # or .env.example for a minimal set

DATABASE_PASSWORD is the only variable Compose refuses to start without. DJANGO_SECRET_KEY is enforced separately by Django at runtime, so set both:

python -c "from django.core.management.utils import get_random_secret_key; print(get_random_secret_key())"

Then:

docker compose up -d                                  # web, db, redis, worker, beat
docker compose --profile with-nginx up -d             # adds nginx on 80/443
docker compose --profile monitoring up -d             # adds prometheus, grafana, exporters

Place model weights before running inference. See Model weights; the stack starts fine without them.

docker compose exec web python manage.py createsuperuser

Check it came up:

http://localhost:8000/health/     {"status": "ok", "version": "2.0.0", "database": "connected"}
http://localhost:8000/admin/
http://localhost:8000/docs/       Swagger UI (schema requires authentication)

For TLS in development, bash scripts/generate-ssl.sh writes a self-signed pair to ssl/. In production put real certificates at ssl/cert.pem and ssl/key.pem.

Local install without Docker

python -m venv .venv
source .venv/bin/activate            # Windows: .venv\Scripts\activate
pip install -r requirements.txt

torch and torchvision are deliberately not in requirements.txt, because the Docker image inherits them from its base image. A local install needs them explicitly, matched to your CUDA version:

pip install torch==2.1.0 torchvision==0.16.0 --index-url https://download.pytorch.org/whl/cu118

Then:

export DJANGO_ENVIRONMENT=development
export DJANGO_SECRET_KEY=dev-secret-key-not-for-production
export GEODJANGO_ENABLED=False        # see below

python manage.py migrate
python manage.py createsuperuser
python manage.py runserver

GEODJANGO_ENABLED is a real decision, not boilerplate. It defaults to True, which loads django.contrib.gis and requires GDAL and GEOS on the system. Without them, every management command fails with Could not find the GDAL library. Setting it to False swaps Projects.field_polygon from a PostGIS PolygonField to a plain JSONField and drops the spatial database backend, so a database created one way is not compatible with the other. Use False for local work without GDAL installed; use True in production against PostGIS.

Celery needs Redis even in development:

docker run -d -p 6379:6379 redis:7-alpine
celery -A agrisynthia worker -l info -Q detection,mapping,celery

The -Q flag is required. See Background tasks.

Model weights

Weights are recorded in the database, not discovered on disk. detection.models.ModelVersion holds the path, version, active flag, and SHA-256 for each fruit type, and inference resolves the path from that row.

Place all six files, including agac.pt for tree counting, in models/:

models/mandalina.pt  models/elma.pt  models/armut.pt
models/seftale.pt    models/nar.pt   models/agac.pt

Then run, in this order:

python manage.py migrate                          # seeds one ModelVersion row per fruit type
python manage.py migrate_model_files              # moves flat .pt into models/<fruit>/v1/weights.pt
python manage.py verify_model_checksums --store   # records the SHA-256 of each file

All three steps are required. The seed migration always records the versioned path models/<fruit>/v1/weights.pt, so weights left in the flat layout are never found and inference fails with Model weights not found. migrate_model_files performs that move and is safe to re-run.

MODEL_CHECKSUM_VERIFY defaults to on outside development. Weights whose SHA-256 does not match the stored value are refused, and so are weights with no stored checksum, since there is nothing to verify against. That is why --store is not optional. To load without verification, set MODEL_CHECKSUM_VERIFY=False.

Activating a new ModelVersion evicts the in-process model cache automatically, so a version swap takes effect without a restart.

Configuration

Every setting comes from environment variables, read from .env in the project root. .env.example is the complete, authoritative list and is kept in sync with the code. deploy/env.production.template is the same set annotated for a production host.

The variables that block startup or encode a decision:

Variable	Default	Notes
DJANGO_SECRET_KEY	none	Required outside development. Startup fails without it.
DJANGO_ENVIRONMENT	development	production enables SSL redirect, HSTS, secure cookies, and checksum verification.
DJANGO_ALLOWED_HOSTS	localhost,127.0.0.1	Required in production.
DATABASE_NAME / _USER / _PASSWORD / _HOST / _PORT	none	There is no DATABASE_URL. If any of name, user, password, or host is missing the app silently falls back to SQLite, so verify the engine after deploying.
GEODJANGO_ENABLED	True	Changes the field_polygon column type and the database backend. See above.
CELERY_BROKER_URL / CELERY_RESULT_BACKEND / REDIS_CACHE_URL	localhost	There is no REDIS_URL. Broker and cache use separate logical databases.
USE_R2	False	Switches media storage to Cloudflare R2 with presigned URLs.
MODEL_CHECKSUM_VERIFY	on outside dev	Refuses unverified weights.
EMAIL_HOST_USER / EMAIL_HOST_PASSWORD	none	Without them, verification, password reset, and 2FA mail fail and only a warning is logged.
ANTHROPIC_API_KEY	none	Powers the marketing site chatbot. Blank disables the endpoint.
SENTRY_DSN	none	Blank disables error reporting.
ODM_ENABLED	True	Set False to skip NodeODM and use pre-processed orthophotos.
RESULT_RETENTION_DAYS	30	Age at which detection results and their media are deleted.
GRAFANA_ADMIN_PASSWORD	agrisynthia	Change it when running the monitoring profile.

Verify the database engine is what you intended:

python manage.py shell -c "from django.db import connection; print(connection.settings_dict['ENGINE'])"

Services

Service	Port	Profile	Description
web	127.0.0.1:8000	default	Django and Gunicorn
db	127.0.0.1:5432	default	PostgreSQL 15 with PostGIS
redis	127.0.0.1:6379	default	Cache, Celery broker, rate limiting
celery_worker	none	default	Task worker
celery_beat	none	default	Periodic scheduler
nginx	80, 443	with-nginx	Reverse proxy and TLS
prometheus	127.0.0.1:9090	monitoring	Metrics
grafana	127.0.0.1:3000	monitoring	Dashboards
redis_exporter	none	monitoring	Redis metrics
postgres_exporter	none	monitoring	Postgres metrics
celery_exporter	none	monitoring	Celery metrics

Everything except nginx binds to loopback, so nothing is reachable from outside the host without the nginx profile.

Background tasks

Celery over Redis, with tasks routed across three queues:

Queue	Tasks
detection	detection.tasks.*
mapping	dron_map.tasks.*
celery	reports.tasks.*, website.tasks.*

A worker started without -Q consumes only the default celery queue. Detection and drone tasks then sit in Redis unconsumed: async detection never leaves PENDING, the progress stream hangs, and the ODM watchdog never runs. Always pass all three:

celery -A agrisynthia worker -l info -Q detection,mapping,celery

Run exactly one celery beat across the fleet. Two schedulers double every periodic task.

Scheduled tasks

Task	Interval	Purpose
detection.tasks.check_model_health	24h	Alerts when mean confidence drops below 0.7
detection.tasks.cleanup_old_results	24h	Deletes results older than RESULT_RETENTION_DAYS
dron_map.watchdog_stuck_odm_tasks	1h	Reaps hung NodeODM jobs
dron_map.refresh_all_sentinel2_ndvi	7d	Pulls new Sentinel-2 scenes
dron_map.send_ndvi_stress_alerts	7d	Emails fields below the stress threshold
reports.tasks.send_scheduled_reports	1h	Dispatches due report schedules
website.tasks.refresh_ndvi_hero	30d	Re-renders the homepage NDVI tile

Manual invocation:

docker compose exec celery_worker celery -A agrisynthia call detection.tasks.check_model_health
docker compose exec celery_worker celery -A agrisynthia inspect active_queues

API

Full schema at /docs/ (Swagger) and /redoc/. Every API endpoint requires authentication via session cookie; only /health/ is public. Anonymous callers are throttled to 10 requests per hour, authenticated to 100.

Detections /api/detections/

Method	Path	Description
GET POST	/api/detections/	List and create
GET	/api/detections/<id>/	Retrieve
GET	/api/detections/statistics/	Aggregates
GET	/api/detections/recent/	Last 10
GET	/api/batches/	Batch detections
GET	/api/batches/<id>/summary/	Batch summary

Projects and analysis /api/projects/

Method	Path	Description
GET POST	/api/projects/	List and create
GET	/api/projects/<id>/summary/	Project summary
GET	/api/projects/<id>/density/	Tree density grid, GeoJSON
GET	/api/projects/<id>/stress-zones/	NDVI stress zones, GeoJSON
GET	/api/projects/<id>/decisions/	Ranked agronomic recommendations
GET	/api/projects/<id>/yield/	Yield estimate
GET	/api/projects/<id>/full-analysis/	All of the above in one call
GET	/api/projects/statistics/	Aggregates
GET	/api/projects/by_farm/ by_state/	Grouped listings

Detection UI /detection/

Method	Path	Description
GET POST	/detection/	Single image detection
GET POST	/detection/mcti/	Multi-image batch
POST	/detection/async-detection/	Queue a job, returns a task id
GET	/detection/task-status/<id>/	Poll job status
GET	/detection/task-stream/<id>/	Server-Sent Events progress stream
GET	/detection/media/<path>	Authenticated media proxy
GET	/detection/system-monitoring/	Resource dashboard
POST	/detection/cache/invalidate/	Clear prediction cache, staff only

Drone mapping /dron-map/: dashboard/, projects/, map/<id>/, overview/, yield/, projects/<id>/ndvi/, projects/<id>/odm-status/

Reports /reports/: list, request/detection/, request/drone/, download/<id>/, delete/<id>/, schedule/create/, schedule/<id>/toggle/, schedule/<id>/delete/

Accounts /accounts/: login/, logout/, register/, profile/, settings/, password-reset/, verify-email/<token>/, 2fa/setup/, verify-2fa/

Public site: / and /en/, with translated slugs (/urun/ and /en/product/). Also robots.txt, sitemap.xml, /health/, /metrics/.

Layout

Path	Role
agrisynthia/	Settings, URLs, Celery app, vegetation indices, inference entry point
detection/	Detection models, views, tasks, and vendored YOLOv7 under detection/yolo/
dron_map/	Drone projects, ODM integration, Sentinel-2 ingest
accounts/	Registration, email verification, TOTP 2FA
reports/	PDF and Excel generation, scheduled email delivery
website/	Public marketing site, blog, chatbot
spatial_analysis/	Density grids and stress zones, plain Python
decision_engine/	Stress zones to recommendations, plain Python
yield_prediction/	Yield scoring, plain Python
analysis_logger/	Persists combined analysis runs, plain Python
deploy/	systemd units and production runbook
monitoring/	Prometheus config and Grafana dashboards

The last four Python packages are not Django apps and are not in INSTALLED_APPS.

Testing

pytest                                    # 165 tests
pytest detection/tests.py                 # one file
pytest -k "stress_zones"                  # by name
python manage.py test detection dron_map  # Django runner, 58 tests

Use pytest for full coverage. spatial_analysis, decision_engine, yield_prediction, and analysis_logger are not Django apps, so the Django test runner never discovers their tests.

Tests use agrisynthia.test_settings, which mocks Redis and runs Celery eagerly. They do not need a running broker. They do inherit the database configuration, so with GEODJANGO_ENABLED=True and no PostGIS available, set GEODJANGO_ENABLED=False first.

detection/test_inference.py runs real inference against real weights. Those tests skip when models/mandalina/v1/weights.pt is absent, so a checkout without weights stays green while a machine that can run inference still catches regressions.

Translations

Turkish is the source language. Message ids are Turkish, so a fresh clone renders correctly in Turkish with nothing compiled. English lives at locale/en/LC_MESSAGES/django.po and is complete; the compiled django.mo is committed so production hosts do not need gettext at deploy time.

python manage.py makemessages -l en    # after changing templates or views
python manage.py compilemessages

Both need GNU gettext (choco install gettext on Windows, apt install gettext on Linux). A pre-commit hook recompiles .mo whenever a .po is staged. Enable it once per clone:

git config core.hooksPath hooks

Deployment

deploy/README.md covers a bare metal install at /opt/agrisynthia: systemd units for web, worker, beat, and backups, plus the system packages and verification steps. For containers, use docker-compose.yml at the root. Do not mix the two on one host.

Backups

Postgres dumps to Cloudflare R2 under the backups/postgres/ prefix:

python scripts/backup_postgres.py backup            # dump, gzip, upload
python scripts/backup_postgres.py list              # newest first
python scripts/backup_postgres.py restore           # restore latest
python scripts/backup_postgres.py prune --days 30   # delete older than 30 days

Needs pg_dump and psql on PATH and R2 credentials in .env. Restore drops existing data and prompts for confirmation unless -y is passed. Schedule it with either the systemd timer in deploy/systemd/ or a cron entry, never both.

Logs

docker compose logs -f web
docker compose logs -f celery_worker

Files rotate at 10 MB in logs/django.log and logs/django_errors.log. The detection and dron_map loggers are set to WARNING by default.

Updating

git pull origin main
docker compose build
docker compose up -d
docker compose exec web python manage.py migrate

Security

• Never commit .env. It is gitignored and excluded from the Docker build context.
• All API endpoints require authentication. Only /health/ is public.
• Media is served through an authenticated proxy using X-Accel-Redirect, or presigned R2 URLs when USE_R2 is on. Direct /media/ access is blocked at nginx.
• Model weights are refused unless their SHA-256 matches the recorded value.
• nginx enforces three rate limits: 10 req/s on /api/, 5 req/s on /accounts/ and /admin/, 100 req/s elsewhere, plus 50 concurrent connections per IP. Django adds per-view limits on login, registration, 2FA, password reset, and uploads.
• Uploads are capped at 10 MB for detection images and 100 MB for drone imagery, and validated by magic bytes rather than file extension.
• CSP, HSTS, secure cookies, and SSL redirect activate automatically outside development.

Load testing

pip install locust
export LOCUST_TEST_USERNAME=loadtest LOCUST_TEST_PASSWORD=...
locust -f scripts/locustfile.py --host https://staging.example.com

Three user classes at weights 6 / 2 / 3: browsing, uploading detections while watching the progress stream, and polling the REST API. Never point it at production; it enqueues Celery work on every iteration and holds streaming connections open. Drop a small JPEG at scripts/_sample.jpg first.

License

Proprietary. Copyright (c) 2026 ChronoCoders. All rights reserved. Contact altug@bytus.io for licensing.