████████╗███████╗██████╗ ██████╗  █████╗ ██╗   ██╗██╗ █████╗ 
╚══██╔══╝██╔════╝██╔══██╗██╔══██╗██╔══██╗██║   ██║██║██╔══██╗
   ██║   █████╗  ██████╔╝██████╔╝███████║██║   ██║██║███████║
   ██║   ██╔══╝  ██╔══██╗██╔══██╗██╔══██║╚██╗ ██╔╝██║██╔══██║
   ██║   ███████╗██║  ██║██║  ██║██║  ██║ ╚████╔╝ ██║██║  ██║
   ╚═╝   ╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝  ╚═══╝  ╚═╝╚═╝  ╚═╝

Satellite Imagery Intelligence — Land Classification & Vegetation Health Analysis

Machine Learning Course Project

---

🛰️ Overview

TerraVia is an end-to-end machine learning platform for satellite imagery analysis. Upload a .tif satellite image and get instant AI-powered insights — land use classification across 8+ categories and a real-time vegetation health index — all through a sleek, browser-based interface backed by a deployed deep learning model. https://github.com/Samarthsawant/land-classification-and-vegetation-health/releases/download/v1/terravia_v2.html

Manual land analysis used to take weeks. TerraVia does it in under 3 seconds.

---

🌍 The Problem

Challenge	Scale
Global land cover has changed	77% of Earth's land area affected
Degraded land worldwide	3.2 billion hectares
Manual satellite analysis time	~3 weeks per region
Traditional methods	Expensive, unscalable, expert-dependent

Remote sensing and land monitoring are critical for climate science, agriculture, urban planning, and disaster response — yet analysis at scale remains inaccessible without expensive software and domain expertise. TerraVia solves this.

---

✨ Features

• 🗺️ Land Use Classification — Classifies satellite imagery into 8+ land cover categories
• 🌿 Vegetation Health Index — Real-time NDVI-proxy score with stress classification (Healthy / Moderate Stress / High Stress)
• 📊 Spectral Breakdown — Top-3 class probability scores for every prediction
• 🖼️ Native TIFF Rendering — Custom pure-JS TIFF parser with contrast stretching for 8-bit and 16-bit satellite imagery — no plugins needed
• ⚡ < 3 Second Inference — Powered by a deployed Hugging Face Space backend
• 📥 Export Results — Download full predictions as JSON
• 📱 Fully Responsive — Works on desktop and mobile

---

🧠 Model & Methodology

Architecture

Raw Satellite Image (GeoTIFF)
         ↓
  Preprocessing & Normalization
         ↓
  Convolutional Neural Network (CNN)
  ┌─────────────────────────────┐
  │  Conv Blocks → BatchNorm    │
  │  MaxPooling → Dropout       │
  │  Dense Layers → Softmax     │
  └─────────────────────────────┘
         ↓
  Multi-class Classification + Confidence Scores
         ↓
  Vegetation Health Computation

Training Pipeline

1. Data Ingestion — EuroSAT imagery loaded and validated
2. Preprocessing — Normalization, augmentation (flips, rotations, spectral jitter)
3. CNN Training — Multi-class classification with cross-entropy loss
4. Evaluation — Per-class accuracy, F1, precision, recall on held-out test set
5. Deployment — Exported model served via Hugging Face Spaces API

---

📦 Datasets

Dataset	Images	Coverage	Bands
EuroSAT	~27,000	Europe (Sentinel-2)	13 spectral bands

Dataset spans diverse biomes, seasons, and resolutions — making the model robust to geographic and atmospheric variation.

---

📈 Results

Metric	Score
Overall Accuracy	96.2%
F1 Score (macro)	94.8%
Forest Precision	97.1%
Urban Recall	91.3%

Per-Class Accuracy

Forest        ████████████████████ 97.1%
Water         ███████████████████  95.8%
Agriculture   ██████████████████   93.4%
Shrubland     █████████████████    91.7%
Urban         █████████████████    91.3%
Barren        ████████████████     89.5%
Wetland       ███████████████      87.2%
Snow/Ice      ██████████████████   94.0%

---

🚀 Getting Started

Use the Live App

Visit the deployed TerraVia interface, upload any .tif satellite image, and get instant results.

Make sure the Hugging Face Space backend is in Running state before analyzing.

Run Locally

# follow the download link
https://github.com/Samarthsawant/land-classification-and-vegetation-health/releases/download/v1/terravia_v2.html

# Open the frontend
open terravia_v2.html

Backend (Hugging Face Space)

The model is deployed at:

https://samdoesitbetter-terravia-api.hf.space/predict

Send a POST request with a multipart/form-data body containing your image file under the key file.

import requests

with open("image.tif", "rb") as f:
    response = requests.post(
        "https://samdoesitbetter-terravia-api.hf.space/predict",
        files={"file": f}
    )

print(response.json())
# {
#   "label": "Forest",
#   "confidence": 97.1,
#   "vegetation_health": 84,
#   "all_scores": { "Forest": 97.1, "Agriculture": 1.8, ... }
# }

---

🗂️ Project Structure

terravia/     
├── terravia_v2.html       # Frontend — full single-file app
├── model/
│   ├── train.py           # Training script
│   ├── evaluate.py        # Evaluation & metrics
│   └── model.pt           # Trained weights
├── backend/
│   └── app.py             # FastAPI inference server (HF Space)
├── assets/
│   └── sample_images/     # Test satellite imagery
└── README.md

---

🛠️ Tech Stack

Layer	Technology
Frontend	Vanilla HTML/CSS/JS — zero dependencies
TIFF Rendering	Custom pure-JS parser (8/16-bit, RGB/grayscale)
ML Framework	PyTorch / TensorFlow
Model Serving	Hugging Face Spaces (FastAPI)
Datasets	EuroSAT

---

🔮 Future Work

• Semantic Segmentation — Pixel-level land cover maps instead of image-level classification
• Temporal Analysis — Change detection over time using multi-date image pairs
• Multi-Spectral Fusion — Full 13-band Sentinel-2 processing (currently RGB+NIR)
• Mobile App — React Native client for field surveys with camera integration

---

Machine Learning Course Project — ECE Department

---

Built with 🛰️ and a lot of satellite images