Spectral Spatial Kmeans

The spectral_spatial_kmeans() function performs a hybrid segmentation on a DataCube by first clustering spectral signatures with standard KMeans and then enforcing spatial coherence through Markov Random Field (MRF) regularization. Initially, each pixel’s spectrum is assigned to one of K clusters based solely on spectral distance. Subsequently, a smoothing step refines labels to encourage neighboring pixels to share the same cluster, reducing salt-and-pepper noise and emphasizing spatially contiguous regions.

Example

The following example demonstrates spectral–spatial KMeans on a synthetically generated DataCube.

import wizard
from wizard._utils.example import generate_pattern_stack
from wizard._processing.cluster import spectral_spatial_kmeans
import matplotlib.pyplot as plt

# generate synthetic data with spatial patterns
data = generate_pattern_stack(12, 400, 400, n_circles=5, n_rects=3, n_triangles=1)

# create a DataCube
dc = wizard.DataCube(data)

# apply spectral–spatial KMeans clustering
labels = spectral_spatial_kmeans(dc, n_clusters=4, spatial_radius=2)

# visualize a wavelength slice and the segmentation
fig, axes = plt.subplots(1, 2, figsize=(10, 5))

axes[0].imshow(dc[7])
axes[0].set_title("Original Slice (λ index 7)")
axes[0].set_axis_off()

axes[1].imshow(labels)
axes[1].set_title("Segmented (Spectral–Spatial KMeans)")
axes[1].set_axis_off()

plt.tight_layout()
plt.show()

The result highlights how pure spectral clustering followed by spatial smoothing yields segments that honor both spectral similarity and contiguous spatial patterns.