Unsupervised Classification — ISODATA

Image classification — the big idea. A process of assigning every pixel in an image to one of a pre-defined set of classes. The output is a thematic map (e.g., forest / crop / water / urban).

• Two broad families: unsupervised (this deck) and supervised (next deck).

Unclassified vs. classified. Left: the raw multispectral image with 11 classes in the legend (Meadow, Lawn, Oak, Residential, Dry grass, Swamp, Lake, Salt evap, Comm/Indu, Transportation…). Right: the same scene after classification — each pixel assigned to one of those classes, shown as a flat color.

• A good thematic map has tight spectral separability between classes and an accuracy assessment to back it up (see slide 3).

Three steps of any classification project.

1. Classification scheme — define the classes (and hierarchy) up front.
2. Classification methods — pick the algorithm (ISODATA, K-means, Max Likelihood, etc.).
3. Accuracy assessment — confusion matrix, user’s / producer’s accuracy, kappa.

(Repeat of slide 3 in the original deck.) Same three-step outline — the instructor uses it as a navigation anchor for the rest of the lecture.

Choosing a classification scheme.

• Classes are always a simplification / generalization of the real world.
• Scheme choice depends on the information-extraction objective (are you counting tree crowns, or mapping land cover?).
• You can pick broad/abstract or specific/concrete classes.
• Hierarchical schemes are standard — coarse top level → refined sub-classes.
• Canonical example: Anderson (1976) land-use / land-cover system for RS data. Level I = 9 classes; Level II = 37 subclasses.

Anderson LULC — Level I (1–4) classes.

Anderson LULC — Level I (5–9) classes.

(Step navigation slide.) Same three-step outline — moving from “scheme” to “methods.”

Multispectral (per-pixel) classification.

• Makes use of spectral response patterns — different land-cover types reflect different amounts in each band, creating separable “clusters” in feature space.
• Per-pixel / point classifiers look at one pixel at a time (no spatial context).
• This approach is the heart of remote-sensing land-cover work, despite not using neighborhood information.

Spectral reflectance curves. Different surface features have distinctive EM reflectance “signatures” across wavelength. That’s why band math and classification work at all.

• Vegetation: low red, high NIR.
• Water: low everywhere, especially NIR.
• Soil: moderate across the board, peaks in SWIR.
• Snow: high in visible, drops in SWIR (how we separate snow from clouds).

Supervised vs. unsupervised — side-by-side comparison.

• Classic trade-off: unsupervised is quick and reveals natural groupings; supervised is more accurate when you have good training data.

Unsupervised classification (clustering) — how it differs from supervised.

• User supplies only a few parameters; no training sites.
• Computer finds statistical patterns (spectral clusters) in the data automatically.
• Spectral classes ≠ meaningful categories. The output is “Cluster 7”, not “Corn.”
• Post-labeling is required — the analyst assigns real-world labels to each cluster after the algorithm runs (often via the Recode tool).

ISODATA — Iterative Self-Organizing Data Analysis Technique.

• Assigns each candidate pixel to its closest cluster mean (minimum spectral distance).
• Iterative: repeatedly re-classifies every pixel and recomputes cluster statistics until stable.
• Derived from ERDAS Imagine’s implementation (Field Guide 2002 p. 232).

ISODATA — three parameters the user must set.

• N — max number of clusters. Becomes the maximum possible number of classes.
• T — convergence threshold. The % of pixels whose class values must stay unchanged between iterations to declare convergence (e.g., 95%, 98%).

• M — max iterations. Safety limit so the algorithm halts even if T is never reached.

• Rule of thumb: if ISODATA returns far fewer clusters than N, your data doesn’t support that many; if it maxes out N, bump the parameter up.

Worked example — computing the convergence threshold T.

The slide’s 10-pixel table shows which pixels changed class between iterations:

• Changed: 3 pixels. Unchanged: 7 pixels.
• T (unchanged %) = unchanged ÷ total = 7 / 10 = 70%.
• If the operator set T = 95%, 70% is not yet high enough → ISODATA runs another iteration.

ISODATA procedure, part 1 — the first pass.

1. Pick N arbitrary cluster means to start.
2. Compute the spectral distance between each pixel and every cluster mean.
3. Assign each pixel to its closest mean.
4. After the pass, recompute each cluster’s mean using the actual pixels that ended up in it — the means shift in feature space.
5. Use the new means for the next iteration.

ISODATA procedure, part 2 — iteration and termination.

• Each new iteration: compare every pixel to the updated means, reassign to the closest cluster.
• Stop condition — whichever comes first:
  • Convergence threshold T is reached (few pixels moving classes).
  • Max iterations M is reached.

ISODATA visualization — three snapshots.

• (1) Five arbitrary initial cluster means plotted in 2-D spectral space.
• (2) After the first pass — cluster means have shifted toward the actual pixel densities.
• (3) Second pass — means have stabilized closer to true cluster centers.

Source: ERDAS Field Guide, Figures 6-3/6-4/6-5, pp. 215–216.