Image Classification → Steps

Assembling the Training Data

Training data (also called training samples or signature sets) are the foundation of supervised image classification in remote sensing.
This is where the analyst selects representative examples of each land-cover class—such as water, vegetation, urban, soil, etc.—from the satellite image.

To prepare training data properly, several analytical and interactive steps are used. These help ensure that the classes are well separated and that the classifier receives the correct spectral information.

1. Graphical Representation of Spectral Response Patterns

✔ What it means

For each class (e.g., water, forest, built-up), the training pixels have a spectral signature—a pattern of reflectance values across the image's spectral bands.

This pattern is visualized using:

• Spectral reflectance curves

• Band-by-band scatter plots

• Histograms for each band

✔ Purpose

• To understand how different classes behave in different bands

• To check if the selected training pixels are spectrally consistent

• To identify overlaps between classes (e.g., dark soil and turbid water)

✔ Key terminology

• Spectral profile / spectral signature

• Spectral separability

• Spectral scatterplot

• Feature space

2. Quantitative Expressions of Category Separation

This step uses mathematical measures to check if classes are well-separated in spectral space.

✔ Why it matters

Classification accuracy depends on how distinct one class is from another.
If training classes overlap too much, classification errors will occur.

✔ Common quantitative measures

• Transformed Divergence (TD)

• Jeffries–Matusita Distance (JM)

• Bhattacharyya Distance (BD)

✔ What these values indicate

• Values close to 2.0 (JM scale) → excellent class separability

• Values close to 0.0 → poor separability; classes overlap

• Helps decide whether to:

  • combine classes

  • redefine training samples

  • collect more samples

  • split a mixed class

✔ Key terminology

• Separability index

• Statistical distance

• Cluster separation

• Spectral overlap

3. Self-Classification of the Training Data Set

✔ Concept

Before performing classification on the full image, the classifier is run only on the training pixels themselves.

✔ Purpose

• To check if the algorithm correctly "recognizes" the classes it was trained on.

• If the classifier mislabels the training samples, the training data need to be corrected.

✔ What it reveals

• Misclassified pixels → inaccurate training sets

• Mixed or overlapping classes

• Inconsistencies in attribute statistics (means, variances)

• Too much variability within a class

✔ Key terminology

• Internal accuracy check

• Confusion among training classes

• Spectral homogeneity

4. Interactive Preliminary Classification

✔ What it is

A rough or temporary classification is generated on the image using preliminary training samples.

✔ Purpose

• To visually inspect how the training data behave when applied to the entire image

• To refine training sites

• To identify new sub-classes or remove misidentified ones

✔ What the analyst checks

• Are water bodies correctly classified?

• Are vegetation areas split properly (forest vs cropland)?

• Are built-up areas being confused with dry soil?

✔ Why "interactive"?

The analyst reviews the output and actively adjusts:

• training polygons

• class definitions

• band combinations

• class separability

✔ Key terminology

• Pre-classification map

• Trial classification

• Interactive refinement

5. Representative Subscene Classification

✔ Concept

Instead of classifying the whole image, a small but representative subscene is used.

A subscene:

• contains all major land-cover types

• captures geographic and spectral variability

• is easier to evaluate and test

✔ Purpose

• To test classifier performance on a manageable area

• To refine spectral signatures before final classification

• To avoid wasting processing time on the full image if training data are weak

✔ What it helps detect

• Class confusion in specific regions

• Spectral variability across the scene

• Need for more training samples

• Problems with similar classes (e.g., shallow water vs wet soil)

✔ Key terminology

• Subscene

• Training refinement

• Pilot classification

• Signature validation

Assembling training data for supervised image classification involves:

1. Graphical representation of spectral response patterns – using spectral curves, histograms, and scatter plots to visualize class behavior.

2. Quantitative expressions of category separation – using statistical measures (JM, TD, BD) to evaluate how distinct classes are.

3. Self-classification of training data – testing if the classifier correctly labels its own training samples.

4. Interactive preliminary classification – producing a trial classification to visually refine training sites.

5. Representative subscene classification – testing the classifier on a smaller, diverse image subset to check accuracy and refine signatures.