Hybrid classification and Post-classification smoothing

Hybrid classification is a combined classification approach that uses both supervised and unsupervised classification techniques together.
It is designed to take advantage of the strengths of each method and to overcome their weaknesses.

What Is Hybrid Classification?

Hybrid classification blends:

• Unsupervised classification (e.g., ISODATA, K-means)

• Supervised classification (e.g., Maximum Likelihood, SVM)

✔ Concept

1. First, an unsupervised algorithm groups pixels into spectral clusters without prior knowledge.

2. These clusters are then labeled or merged into meaningful land-cover classes using supervised training data.

✔ Why use hybrid methods?

• Unsupervised classification captures natural spectral groupings.

• Supervised classification improves accuracy by using reference samples.

• Together, they reduce errors caused by poor training data or complex landscapes.

✔ Key Terminology

• Cluster: a group of pixels with similar spectral characteristics.

• Signature training: giving labels to clusters.

• Spectral homogeneity: similarity within a cluster.

• Class merging: combining multiple clusters into one land-cover type.

Hybrid Classification

Step 1: Unsupervised Clustering

Algorithms like ISODATA or K-means group pixels into 20–50 clusters based only on spectral properties.

Step 2: Assign Clusters to Land-Cover Classes

Use training samples, field data, or expert knowledge to assign each cluster to:

• water

• forest

• agriculture

• built-up

• soil
  (or other classes)

Step 3: Supervised Refinement

Run a supervised classifier (e.g., Maximum Likelihood) using the cluster-based signatures.

Step 4: Merge & Edit Classes

Check for:

• confused clusters

• isolated patches

• mixed clusters
  Clusters are merged or adjusted.

Step 5: Final classified image

A clean, corrected classification map is produced.

Advantages of Hybrid Classification

• Improves accuracy in heterogeneous landscapes

• Handles mixed pixels better

• Reduces reliance on perfect training samples

• Captures subtle spectral differences

• Reduces spectral confusion between similar classes

✔ Suitable for:

• complex landscapes

• urban environments

• vegetation mosaics

• large areas with limited training data

Limitations

• More interactive and time-consuming

• Requires expertise for cluster labeling

• Too many clusters can make interpretation difficult

Post-Classification Smoothing

After classification, the resulting land-cover map often has:

• salt-and-pepper noise

• scattered small patches

• isolated mislabeled pixels

Post-classification smoothing removes these artifacts to produce a cleaner, generalized map.

What Is Post-Classification Smoothing?

It is the process of cleaning and refining a classified image by applying spatial filters or majority rules to reduce noise and improve map readability.

✔ Why smoothing is needed?

Because pixel-based classifiers classify each pixel individually, ignoring spatial relationships.
This results in:

• random noisy pixels

• speckled appearances

• unrealistic boundaries

Smoothing creates spatially coherent regions.

Common Smoothing Techniques

A. Majority (Mode) Filter

• A moving window (3×3, 5×5) scans the classified image.

• Each pixel is replaced by the most common class in the window.

• Removes small patches and isolated noise.

B. Median Filter

• Similar to majority filter but uses median instead of majority.

• Preserves edges better.

C. Morphological Operations

• Opening: removes small isolated pixels.

• Closing: fills gaps in homogeneous regions.

D. Region Growing

• Groups contiguous pixels belonging to the same class into larger coherent regions.

E. Elimination of Small Patches

• Removes polygons smaller than a defined threshold (e.g., <1 hectare).

Results of Smoothing

• More realistic class shapes

• Reduced classification noise

• Better readability for maps

• Improved accuracy for urban and natural landscapes

• Cleaner boundaries between classes