🛰️ SentinelChange-AI

Copernicus Sentinel-2 Change Detection with GPU-Accelerated Deep Learning (ChangeStar + SAM + ESRGAN)

High-resolution LEVIR-CD benchmark — ChangeStar fine-tuned baseline

Copernicus Sentinel-2 (10 m) experiment — work in progress, Ground Truth Refinement , Model Re-Training , GPU optimization ongoing

---

🌍 Overview

SentinelChange-AI is an applied research pipeline for multi-temporal satellite change detection using Copernicus Sentinel-2 imagery.
It integrates ChangeStar for coarse semantic change prediction, Segment Anything Model (SAM) for precise boundary refinement, and ESRGAN/RRDBNet for super-resolving 10 m scenes to near-aerial visual quality.

Beyond accuracy, this repository emphasizes GPU acceleration, profiling, and workload optimization — ideal for researchers, practitioners, and AI engineers exploring geospatial deep learning on high-volume remote-sensing data.

---

🔑 Key Features

• 🛰️ Copernicus Sentinel-2 Change Detection Pipeline (10 m resolution)
• ⚙️ ChangeStar Deep Change-Detection Backbone (Zheng et al., CVPR 2022)
• ✂️ SAM-Based Label Refinement for high-quality masks
• 🔍 Automatic Sentinel-2 Pairing by tile ID & year (MGRS)
• 🧠 Super-Resolution (ESRGAN / RRDBNet) for visual enhancement
• 🚀 GPU-Optimized Training & Inference workflows (CUDA / TorchRun)
• 📊 Benchmark-grade workloads for Nsight profiling and GPU efficiency studies

---

🧭 Attribution

This repository builds upon the open-source ChangeStar framework by Zheng et al. (CVPR 2022).
Model definitions, configuration templates, and baseline training logic are reused under their MIT license.
All Copernicus Sentinel-2 processing, SAM integration, ESRGAN upscaling, and GPU acceleration workflows were independently developed by Atul Vaish.

---

📂 Repository Structure

SentinelChange-AI/
├── ChangeStar/                       # Forked base framework (Zheng et al.)
│   ├── configs/                      # Model configs for training/fine-tuning
│   ├── module/                       # Core model architectures
│   ├── generate_labels_sam.py        # SAM label generation (1024×1024 tiles)
│   ├── generate_labels_sam_upscaled.py
│   ├── inference_sentinel.py         # Tile-wise inference visualization
│   ├── train_sup_change.py           # Supervised change-detection training
│   ├── runtraining.sh                # TorchRun GPU training launcher
│   └── ...
│
├── pair_image_files.py               # Pair Sentinel-2 SAFE scenes by year/tile
├── process_images.py                 # Convert .SAFE → RGB tiles (10 m)
├── upscale_images.py                 # ESRGAN upscaling to ×4 resolution
├── RRDBNet_arch.py                   # RRDBNet architecture for ESRGAN
├── unzip_dataset.py                  # Extract all .SAFE.zip archives
├── results/
│   ├── levir_cd_1.png                # Benchmark reference result
│   └── copernicus_cd_1.png           # Copernicus 10 m sample result
└── datasets/                         # Input SAFE files & processed tiles

---

🧠 End-to-End Workflow

1️⃣ Unzip Copernicus .SAFE archives

python unzip_dataset.py

2️⃣ Pair multi-temporal Sentinel-2 scenes by tile ID and year

python pair_image_files.py

3️⃣ Convert spectral bands to RGB tiles (10 m)

python process_images.py

4️⃣ (Optional) Enhance tiles using ESRGAN ×4 super-resolution

python upscale_images.py

5️⃣ Generate SAM-refined labels from ChangeStar predictions

python ChangeStar/generate_labels_sam.py
# or for upscaled data
python ChangeStar/generate_labels_sam_upscaled.py

6️⃣ Train or fine-tune ChangeStar on your dataset

bash ChangeStar/runtraining.sh

7️⃣ Visualize predictions and ground truth

python ChangeStar/inference_sentinel.py

---

⚙️ Environment Setup

conda create -n change_detection python=3.9
conda activate change_detection

pip install torch torchvision torchaudio
pip install ever-alpha tqdm opencv-python pillow rasterio matplotlib segment-anything
pip install numpy urllib3

For ESRGAN upscaling, ensure GPU support and install PyTorch ≥ 2.0 with CUDA 11.x or higher.
For ChangeStar training, recommended GPUs: RTX 3060+, A100, or Jetson Orin for edge experiments.

---

🛰️ Datasets

• LEVIR-CD — 30 cm high-resolution aerial imagery dataset (benchmark for validation).
• Copernicus Sentinel-2 — 10 m multispectral imagery (ESA Copernicus Open Access Hub).
  • Automatically paired by MGRS Tile ID and acquisition year.
  • RGB tiles are created using bands B04, B03, B02.
  • SAM-refined masks used as weak supervision for fine-tuning.

Experiments on Copernicus data are in progress; visual quality and segmentation precision improve significantly when super-resolved with ESRGAN and retrained with ChangeStar.

---

🚀 GPU Acceleration & Profiling

Training and inference workloads are ideal for GPU utilization analysis and Nsight profiling.

Example profiling command:

nsys profile -t cuda,nvtx -o runs/profile bash ChangeStar/runtraining.sh

You can track:

• SM efficiency & occupancy
• Tensor Core utilization
• DRAM throughput
• Kernel launch latency
• GPU power draw & timeline traces

This repository forms part of a broader Agentic GPU Optimization research effort — profiling model workloads to enable autonomous sense–think–act–learn optimization loops for AI acceleration.

---

🧩 Research & Applications

• 🌾 Land-use and Land-cover change analysis
• 🏙️ Urban expansion and infrastructure mapping
• 🌊 Flood and disaster impact assessment
• 🌋 Environmental monitoring (deforestation, mining, glaciers)
• 🛰️ GPU benchmarking for Earth Observation AI workloads

---

🧱 Roadmap

Milestone	Description	Status
✅ Baseline Copernicus Change Detection	End-to-end pipeline completed	Done
✅ ESRGAN Super-Resolution	Working upscaler integrated	In Progress
🧩 SAM Integration	Precise label refinement	In Progress
🧠 DINOv2 + SAM Fusion	Improved embedding quality	In Progress
⚙️ GPU Profiling Integration	Nsight metrics + agentic optimization	In Progress
🌐 Multi-sensor Fusion	Extend to Sentinel-1 & Landsat-8	Planned

---

🏗️ Work in Progress

• 🌐 Fusion of DINOv2 + SAM for adaptive feature extraction
• 🔬 Nsight-driven GPU workload profiling module
• 🧩 Integration with agentic GPU optimization (Sense → Think → Act → Learn)
• 🛰️ Expansion to Sentinel-1/3 multimodal change detection

---

📄 License

This project adopts the MIT License, consistent with the original ChangeStar.
All additional modules authored by Atul Vaish are released under the same terms.

---

✨ Citation

If you use or reference this work, please cite both the original ChangeStar paper and this applied project:

@inproceedings{zheng2022changestar,
  title={ChangeStar: A Universal Change Detection Framework},
  author={Zheng, Zilong and Zhong, Yanfei and Wang, Zheng and Li, Zhen and Zhang, Liangpei},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

@software{vaish2025_sentinelchangeai,
  author = {Atul Vaish},
  title  = {SentinelChange-AI: Copernicus Sentinel-2 Change Detection with GPU-Accelerated Deep Learning},
  year   = {2025},
  url    = {https://github.com/avaish/SentinelChange-AI}
}

---

👨‍💻 Author

Atul Vaish
Independent Applied AI & Geospatial Research
GPU Optimization | Satellite AI | Edge Intelligence
🌐 https://aifusion.in
📧 atul7911@gmail.com
📍 India | EU Collaboration Ready