AeroScan

AeroScan is a web app for scanning commercial roof footprints, scoring rainwater-harvesting potential, and enriching leads with satellite imagery, ML (cooling-tower detection), and AI-generated briefs. It was built for a hackathon-style workflow: map-first “command center,” saved buildings, PDF reports, and optional voice playback.

---

Features

• Interactive map — Leaflet-based view with Earth Engine raster layers where configured.
• Viewport scan — Pulls building roof footprints from Overture Maps (DuckDB + GeoParquet on S3) for CONUS, or uses Google Open Buildings via Earth Engine where applicable; merges nearby polygons into site-scale leads.
• Lead scoring — Heuristic tier, viability, and component scores using roof area, CHIRPS rainfall samples, and simple cooling-tower priors.
• ML — YOLOv5-seg + EfficientNet-B5 pipeline on NAIP (US) or Sentinel-2 tiles for cooling-tower-style detection (/api/cv/detect-tower).
• AI — Google Gemini for per-building insights; ElevenLabs TTS reads the on-screen insight (no duplicate Gemini call for playback).
• Reports — ReportLab PDF generation with charts and metrics.
• Auth — Auth0 (Next.js middleware) for sign-in.
• Persistence — Postgres/PostGIS for saved leads (Docker Compose).

---

Stack

Layer	Technology
Frontend	Next.js 16, React 19, TypeScript, Tailwind CSS 4, Radix UI, Leaflet / react-leaflet, Recharts
Backend	Python 3.10+, FastAPI, Uvicorn
Geo / data	Google Earth Engine API, DuckDB + spatial/httpfs, Overture Maps release GeoParquet
ML / CV	PyTorch, torchvision, OpenCV (headless), EfficientNet (efficientnet-pytorch), YOLOv5 (via cached torch.hub ultralytics repo), seaborn (hub dependency)
AI / voice	google-generativeai (Gemini), ElevenLabs REST API
PDF	ReportLab
Database	PostgreSQL + PostGIS (optional local via Docker)
Auth	Auth0 Next.js SDK v4

---

Repository layout

SMUHackathon/
├── backend/           # FastAPI app (server.py, cv_service, EE helpers, PDF, Gemini, etc.)
├── frontend/          # Next.js UI (app router, components, Auth0)
├── cv/                # Shared tile fetch / CV utilities used by the API
├── data/              # Stores, ingestion scripts, Overture shard index JSON
├── docker/            # Postgres init scripts for Compose
├── docker-compose.yml # Local PostGIS
├── requirements.txt   # API + app Python deps (install torch/opencv for CV — see below)
├── requirements-cv.txt# Explicit CV stack (torch, opencv, ultralytics, …)
└── *.pt               # Trained weights (YOLO seg + B5) — keep in repo root or set YOLO_WEIGHTS_PATH

---

Prerequisites

• Node.js 20+ (or current LTS compatible with Next 16)
• Python 3.10+ (3.11+ recommended for ML wheels)
• Docker Desktop (optional, for Postgres)
• Accounts / credentials (as needed for your workflow):
  • Google Cloud project with Earth Engine enabled + local gcloud / earthengine authenticate
  • Gemini API key
  • ElevenLabs API key (TTS)
  • Auth0 application (Regular Web App)
• Weights in repo root (defaults expected by backend/cv_service.py):
  • cooling_tower_yolov5seg_best.pt
  • b5_unweighted_best.pt

---

Quick start

1. Clone and environment files

cp .env.example .env
# Edit .env: DATABASE_URL, GEE_PROJECT_ID / GOOGLE_CLOUD_PROJECT, GEMINI_API_KEY,
# ELEVENLABS_API_KEY, Auth0 vars, etc.

For the Next app, mirror Auth0 and NEXT_PUBLIC_API_URL in frontend/.env.local if you do not load them from the repo root (Auth0 SDK typically reads env from the Next directory).

2. Database (optional but recommended for “Saved” leads)

docker compose up -d
# Default DATABASE_URL uses host port 5433 — see .env.example

3. Python API

python3 -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -r requirements.txt
# CV endpoints also need PyTorch + OpenCV; if not already pulled in:
pip install -r requirements-cv.txt

Start the API from the repository root with PYTHONPATH set:

set -a && source .env && set +a   # bash/zsh
PYTHONPATH=. uvicorn backend.server:app --reload --host 0.0.0.0 --port 8000

4. Frontend

cd frontend
npm install
npm run dev

Open http://localhost:3000. The UI defaults to http://localhost:8000 for the API unless NEXT_PUBLIC_API_URL is set.

5. Production-style frontend

cd frontend
npm run build
npm run start

---

Configuration highlights

Full reference: .env.example. Notable variables:

Variable	Purpose
DATABASE_URL	Postgres connection (saved buildings / scans)
GEE_PROJECT_ID / GOOGLE_CLOUD_PROJECT	Earth Engine project
GEMINI_API_KEY	Building insights
ELEVENLABS_API_KEY	Lead-brief TTS
AUTH0_*, APP_BASE_URL	Auth0 + callbacks
NEXT_PUBLIC_API_URL	Browser → API base URL
YOLO_WEIGHTS_PATH	Override default YOLO checkpoint path
CV_DETECT_TIMEOUT_S	Cap for long-running CV requests
OVERTURE_NAMES_PRIMARY_ONLY	DuckDB fallback if names struct differs on a shard

---

How it fits together

flowchart LR
  subgraph client [Browser]
    UI[Next.js + Leaflet]
  end
  subgraph api [FastAPI]
    EE[Earth Engine layers / CHIRPS sample]
    OV[Overture via DuckDB]
    CV[CV: NAIP/S2 + YOLO + B5]
    GM[Gemini]
    PDF[ReportLab PDF]
    TTS[ElevenLabs TTS]
  end
  DB[(Postgres)]
  UI --> api
  api --> DB
  OV --> S3[(Overture S3 Parquet)]

Loading

---

Scripts and tooling

• Lint (frontend): cd frontend && npm run lint
• Overture shard index: If data/overture_building_parts_<release>.json is missing, large scans may fall back to a slower full S3 glob; see comments in backend/overture_buildings.py and any scripts/ helpers for refreshing the index.

---

Troubleshooting

• No module named 'seaborn' when running CV — install from requirements.txt (pip install seaborn).
• First CV request is slow — Earth Engine auth, model load, and tile fetch; later calls are faster.
• Auth0 errors locally — Ensure callback/logout URLs match APP_BASE_URL and that secrets live where Next can read them (frontend/.env.local).
• Apple Silicon + Docker Postgres — docker-compose.yml pins the PostGIS image to linux/amd64 for compatibility; expect emulation.

---

License

This project was developed for a hackathon. Add or replace this section with your chosen license before public distribution.