Object-Based Image Segmentation using the Felzenszwalb Algorithm

This repository demonstrates how object-based image segmentation can be performed in Python using the Felzenszwalb graph-based algorithm, replicating the multiresolution segmentation approach implemented in eCognition.
The workflow applies to a Sentinel-2A subset and aims to delineate meaningful land-cover objects such as vegetation, water, and built-up areas.

---

Overview

Traditional pixel-based classification methods often ignore spatial context.
Object-Based Image Analysis (OBIA) overcomes this by grouping pixels into homogeneous objects that represent real-world features.
In this notebook, segmentation is achieved through the Felzenszwalb–Huttenlocher algorithm (2004), implemented in scikit-image. The results are compared against eCognition’s Multiresolution Segmentation using identical data and similar parameters.

---

Workflow Summary

1. Load Sentinel-2A subset (6 bands)
   • Bands: Blue, Green, Red, NIR, MIR1, MIR2
2. Apply layer weights
   • (B, G, R, NIR, MIR1, MIR2) = (1, 1, 1, 3, 3, 3)
3. Normalize each band
   • Per-band 2–98% percentile stretch → 0–1 scale
4. Run segmentation
   • Algorithm: Felzenszwalb (skimage.segmentation)
   • Parameters:
     • scale = 180
     • sigma = 1.2
     • min_size = 3200
5. Visualize results
   • False-color composite (NIR–R–G)
   • Overlay boundaries using mark_boundaries()
6. Compare with eCognition
   • eCognition total objects: 67
   • Python total objects: 98

---

Key Libraries

• rasterio – Reading and handling GeoTIFF rasters
• numpy – Array operations and mathematical processing
• matplotlib – Visualization and plotting
• geopandas – Vector data management (optional export step)
• scikit-image – Image segmentation (Felzenszwalb, mark_boundaries)

---

Reference

Felzenszwalb, P. F., & Huttenlocher, D. P. (2004).
Efficient graph-based image segmentation.
International Journal of Computer Vision, 59(2), 167–181.