yantrikaran-innovations/segformer-b5-andhra-landuse

SegFormer-B5 Andhra Pradesh Land Use & Change Detection Model

---

📌 Model Overview

This model is a SegFormer-B5 architecture fine-tuned for semantic land use segmentation and bi-temporal change detection on satellite imagery of 5 districts in Andhra Pradesh, India. It was developed by Yantrikaran Innovations Pvt. Ltd. as part of a smart property identification system commissioned for Andhra Pradesh government land analytics.

The model can:

1. Segment land use from a single satellite tile into 5 classes.
2. Detect changes between two temporal satellite images (e.g., new construction, encroachment, deforestation).

---

🏷️ Label Classes

ID	Class	Description
0	Background	Sky, shadows, uncategorized
1	Buildings	Residential & commercial structures
2	Road	Roads, highways, pathways
3	Water Body	Rivers, lakes, ponds, irrigation canals
4	Open Plot	Vacant land, agricultural/open land

---

📊 Training Metrics

Training Configuration

Parameter	Value
Architecture	SegFormer-B5
Dataset	20,000+ chips, Andhra Pradesh (5 Districts)
Image Source	Sentinel-2 Satellite Imagery
Chip Size	512×512 px
Total Epochs	60 (Best at Ep 56)
Early Stopping Patience	10 epochs

Final Performance (Best Checkpoint - Epoch 56)

Metric	Value
Training Loss	0.8196
Validation Loss	0.8482
Best mIoU	0.5474
Learning Rate (final)	4.04e-06
Training Time	~2698s / epoch

Per-Class IoU (at best checkpoint)

Class	IoU
Background	0.880
Water Body	0.635
Building	0.430
Road	0.431
Open Plot	0.281

---

🚀 Quick Start — Land Use Segmentation

import requests

API_URL = "https://huggingface.co/yantrikaran-innovations/segformer-b5-andhra-landuse"

# If hosted as a Gradio/FastAPI Space:
SPACE_URL = "https://yantrikaran-innovations-segformer-b5-andhra.hf.space/predict"

with open("your_satellite_tile.jpg", "rb") as f:
    response = requests.post(SPACE_URL, files={"file": f})

if response.status_code == 200:
    with open("landuse_mask.png", "wb") as f_out:
        f_out.write(response.content)
    print("✅ Land use mask saved!")
else:
    print(f"❌ Error: {response.text}")

---

🔄 Quick Start — Change Detection

import requests

SPACE_URL = "https://yantrikaran-innovations-segformer-b5-andhra.hf.space/detect"

with open("past_image.jpg", "rb") as f_past, open("new_image.jpg", "rb") as f_new:
    files = {
        "image_past": f_past,
        "image_present": f_new
    }
    data = {
        "threshold": 0.85  # Sensitivity: lower = detect more changes
    }
    response = requests.post(SPACE_URL, files=files, data=data)

if response.status_code == 200:
    with open("change_detection_map.png", "wb") as f_out:
        f_out.write(response.content)
    print("✅ Change map saved!")
else:
    print(f"❌ Error: {response.text}")

---

🗺️ Dataset Details

• Geography: 5 districts of Andhra Pradesh, India
• Data Source: Sentinel-2 MSI (Multispectral Instrument)
• Total Training Chips: 20,000+
• Chip Size: 512×512 pixels
• Annotation Method: Manual + semi-automated QGIS labeling pipeline
• Classes: Water Body, Road, Buildings, Open Plots, Background

---

🏢 About Yantrikaran Innovations

Yantrikaran Innovations Pvt. Ltd. is an AI & robotics company specializing in:

• AI-powered terrain intelligence & geospatial analysis
• Autonomous systems & drone-based recce verification
• Defence technology & decision-support systems for field operations
• IoT and autonomous underwater vehicles (AUVs)

🌐 https://yantrikaran.com
🤗 Hugging Face Organization

---

⚖️ License

This model is released under the Apache 2.0 License.
For commercial or government deployment inquiries, contact: info@yantrikaran.com

---

📝 Citation

If you use this model in your research or application, please cite:

@misc{yantrikaran2025segformer,
  author       = {Yantrikaran Innovations Pvt. Ltd.},
  title        = {SegFormer-B5 Andhra Pradesh Land Use Classification & Change Detection},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/yantrikaran-innovations/segformer-b5-andhra-landuse}},
}