wenmin-wu/cv-spectral-index-segmentation

Spectral Index Segmentation

Overview

Multispectral satellites capture bands beyond visible RGB (NIR, red-edge, SWIR). Normalized difference indices combine these bands into single-channel images where target features (water, vegetation, chlorophyll) have distinct value ranges. A simple threshold on the index produces a binary segmentation mask without any ML model.

Quick Start

import numpy as np
from skimage.transform import resize

def ndwi(green, nir):
    return (green - nir) / (green + nir + 1e-10)

def ccci(red, red_edge, mir):
    return ((mir - red_edge) / (mir + red_edge + 1e-10) *
            (mir - red) / (mir + red + 1e-10))

def ndvi(red, nir):
    return (nir - red) / (nir + red + 1e-10)

ms = load_16band_image(image_id)
re = resize(ms[5], rgb.shape[:2])
mir = resize(ms[7], rgb.shape[:2])
r = rgb[:, :, 0].astype(float)

index = ccci(r, re, mir)
water_mask = (index > 0.11).astype(np.uint8)

Workflow

1. Load multispectral bands and resize to a common spatial resolution
2. Select bands for the target index (e.g., green+NIR for NDWI)
3. Compute the normalized ratio — values typically range [-1, 1]
4. Threshold to produce a binary mask
5. Optionally apply morphological cleaning (opening/closing)

Key Decisions

• Index choice: NDWI for water, NDVI for vegetation, CCCI for chlorophyll/waterways
• Threshold tuning: varies by sensor, scene, and season — tune on labeled samples
• Band alignment: always resize lower-res bands to match before computing indices
• vs ML models: spectral indices are interpretable baselines; often competitive for simple targets

References

• Waterway 0.095 LB