Flashcards

10 cards showing.

Supervised vs. unsupervised classification — key differences?

essential classification

Two camps of pixel classification:

👨‍🏫 Supervised — needs a teacher
- Algorithm: Maximum Likelihood (most common)
- A priori knowledge: required — you give it training samples
- Control: user-driven — you set the class scheme
- Strength: more accurate when training is good
🤖 Unsupervised — clusters on its own
- Algorithm: ISODATA (or K-means)
- A priori knowledge: not required — no training
- Control: computer-automated — it finds the groupings
- Strength: reveals natural groupings, fast first pass

After unsupervised you still label the clusters by hand — that’s where Recode comes in.

💡

Supervised needs a TEACHER (training samples). Unsupervised is CLUSTERING — the computer invents the classes, you label afterward.

classification/sup-vs-unsup

ISODATA — three parameters the user must set?

essential classification

N — max number of clusters (= max classes).
T — convergence threshold: % of pixels unchanged between iterations to declare convergence.
M — max iterations (safety cap).

💡

N-T-M: Number (clusters), Threshold (convergence %), Max-iterations. Terminates on whichever hits first: T reached or M hit.

classification/isodata-params

Maximum Likelihood classifier — how it works and its main assumption?

essential classification

Picture each land-cover class (forest, water, city) as a fuzzy cloud floating in spectral space. Max Likelihood asks the question:

“Which cloud is this pixel most likely sitting inside?”

Crucially, it considers each cloud’s shape (covariance), not just its center. Min Distance only looks at the center — it’s like deciding which city you’re closest to without checking which one’s borders you’re inside.

That’s why ML usually wins on accuracy: it understands a tightly-grouped class is more confident, while a wide loose class lets in more variety.

🔬 Science / formula

For each pixel, compute the probability it belongs to each class. Assign to the most likely class.

Key assumption: each class is multivariate Gaussian (each band normally distributed inside the class).
Uses the class mean AND covariance matrix — that’s how it accounts for class shape, not just position.
Strength: most accurate of the four decision rules when classes are well-sampled and bands are roughly normal.
Weakness: needs enough training samples to estimate covariance; fails on non-Gaussian classes.
Bayesian variant (Hord 1982): supply per-class prior probabilities instead of assuming equal priors.

💡

ML uses **mean + covariance**, not just the mean. Min Distance ignores shape (just the mean). Mahalanobis adds shape. Max Likelihood adds priors on top. The full discriminant formula lives in the long-form review — the *concept* lives here.

classification/ml-discriminant

Supervised classification — three-step procedure?

essential classification

Select training samples — homogeneous AOIs per class.
Generate & evaluate statistical signatures — mean, std, covariance per band.
Class assignment via a decision rule (Min Distance, Max Likelihood, etc.).

💡

Train → Stats → Classify. You TEACH the computer with samples, it LEARNS signatures (mean + covariance), then APPLIES a decision rule to every pixel.

classification/supervised-training

Four common decision rules for supervised classification?

likely classification

The four ways the computer assigns a pixel to a class, ranked by sophistication:

📦 Parallelepiped
- How: box-shaped regions defined by each class’s min/max in every band
- Pro: simple, fast first pass
- Con: leaves gaps + has corner overlap problems
📏 Minimum Distance
- How: assign to whichever class mean is closest (Euclidean)
- Pro: no unclassified pixels, very fast
- Con: ignores class shape, force-fits outliers
📐 Mahalanobis Distance
- How: Min Distance scaled by each class’s covariance
- Pro: accounts for class shape (ellipsoidal)
- Con: still assumes equal priors
🎯 Maximum Likelihood
- How: pick the most probable class, assuming Gaussian distributions
- Pro: most accurate when classes are well-sampled
- Con: needs lots of training data, fails on non-Gaussian classes

classification/four-decision-rules

Minimum Distance classifier — advantages and disadvantages?

likely classification

✅ Advantages - ⚡ No unclassified pixels (every pixel has some nearest mean) - 🚀 Very fast decision rule

❌ Disadvantages - 🎯 Force-fits outlier pixels that should be flagged unclassified - 📐 Ignores class variability — treats tight clusters and loose clusters equally

💡

Min Distance: everyone gets a class (no gaps) but weird pixels get force-fit. Ignores shape. Opposite of Parallelepiped (which leaves gaps).

classification/min-distance-tradeoffs

Parallelepiped classifier — advantages and disadvantages?

likely classification

✅ Advantages - ⚡ Very simple, very fast - 🎯 Good first-pass broad classification

❌ Disadvantages - 🕳️ Gap regions — pixels outside all boxes stay unclassified - ↗️ Corner overlaps — pixels in two boxes get assigned ambiguously

💡

Boxes. Pixel in a box → class. Pixel in NO box → unclassified (gap). Pixel in 2+ boxes (corner overlap) → wrong class. Fast but sloppy.

classification/parallelepiped-tradeoffs

Anderson (1976) LULC — nine Level I classes?

likely classification

🏙️ 1 — Urban / Built-up
🌾 2 — Agricultural
🌿 3 — Rangeland
🌲 4 — Forest
💧 5 — Water
🪷 6 — Wetland
🪨 7 — Barren Land
❄️ 8 — Tundra
🧊 9 — Perennial Snow / Ice

📚 Level II adds 37 subclasses total — used as the standard hierarchical scheme for RS data.

classification/anderson-level1

Covariance between two bands — formula?

maybe classification

\[C_{QR} = \frac{\sum_{i=1}^{k}(Q_i - \bar Q)(R_i - \bar R)}{k - 1}\]

Measures how Band Q and Band R vary together around their means. The full covariance matrix is what Max Likelihood and Mahalanobis use.

classification/covariance-formula

ISODATA convergence threshold T — worked example (10 pixels, 3 changed)?

maybe classification

changed = 3
unchanged = 7
T = unchanged ÷ total = 7/10 = 70%

If user set T = 95%, 70% is not enough → run another iteration.

classification/isodata-convergence-example

Flashcards

changed = 3

unchanged = 7