Sentinel-CD: Satellite Change Detection

Given two co-registered images of the same area taken at different times, Sentinel-CD produces a binary mask highlighting regions that have changed — primarily building construction and demolition.

---

Demo — Hitech City, Hyderabad (2016 → 2026)

Before (2016)	After (2026)

Detected changes (red):

New IT campuses and residential complexes constructed between 2016 and 2026 in Hitech City, Hyderabad.

---

UI

Empty state

Detection results

AI Analysis

---

Results

Metric	Score
F1	0.77
IoU	0.63
OA	0.98

Trained for 50 epochs on LEVIR-CD 256×256 with a Siamese ResNet34-UNet backbone.

---

Project structure

sentinel-cd/
├── model/
│   ├── model.py               # SiameseUNet architecture
│   └── best_model.pth         # trained checkpoint (not tracked by git)
├── inference/
│   ├── sliding_window.py      # inference on large images
│   └── visualize.py           # overlay mask on image
├── api/
│   └── app.py                 # FastAPI REST endpoint
├── assets/
│   ├── hitech_city_2016.png   # demo T1 image
│   ├── hitech_city_2026.png   # demo T2 image
│   └── change_overlay.png     # demo result
├── notebooks/
│   └── levir_cd_v2.ipynb      # Kaggle training notebook
└── requirements.txt

---

Setup

git clone https://github.com/keya115251/sentinel-cd.git
cd sentinel-cd
pip install -r requirements.txt

Download your trained checkpoint from Kaggle and place it at model/best_model.pth.

---

Sliding window inference

python inference/sliding_window.py \
    --t1 path/to/before.png \
    --t2 path/to/after.png \
    --checkpoint model/best_model.pth \
    --output output/change_mask.png

Option	Default	Description
--stride	128	Overlap between patches (128 = 50%)
--threshold	0.5	Change probability threshold

---

Visualize results

python inference/visualize.py \
    --t2 path/to/after.png \
    --mask output/change_mask.png \
    --output output/change_overlay.png

---

REST API

uvicorn api.app:app --host 0.0.0.0 --port 8000

curl -X POST http://localhost:8000/predict \
  -F "t1=@before.png" \
  -F "t2=@after.png" \
  --output change_mask.png

Open http://localhost:8000/docs for the Swagger UI.

---

Training

Open notebooks/levir_cd_v2.ipynb in Kaggle with a T4 GPU and the levir-cd-256 dataset. Run all cells. Expected runtime ~3–4 hours.

---

Model

Architecture	Siamese UNet, ResNet34 encoder
Pretrained	ImageNet
Input	Two RGB patches concatenated → (B, 6, H, W)
Output	Binary change mask → (B, 1, H, W)
Loss	BCE + Dice, pos_weight=10
Dataset	LEVIR-CD 256×256