REMOTE SENSING MIDTERM UNIT 5

Hard vs Fuzzy Classification

Hard Classification: One possible end product of remote sensing imagery is a discrete class for each pixel in an image.

Fuzzy Classification: A similar goal is to determine the % composition of classes within each pixel (since most pixels are mixtures of materials)

output of classification

nominal variable

T/F If two classes have identical reference training state variables, they can not be distinguished using RS data alone

TRUE

Three types of classification

1. Unsupervised

2. Supervised

3. Hybrid

Supervised Classification

Requires “training pixels”, pixels where both the spectral values and the class is known.

Analyst specifies certain “known” areas in the image, and statistics about the DNs in these areas are used to categorize the entire scene

This is referred to as the “training” stage

Unsupervised Classification

No extraneous data is used: classes are determined purely on difference in spectral values.

No training stage, pixels are run through an iterative clustering algorithm, and “similar” groups of pixels are classified thematically

4 Supervised Classification types

Non-Parametric

1. Minimum distance

2. Parallelpiped

3. Nearest-Neighbor Classifiers

Parametric

1. Gaussian Maximum Likelihood

Minimum Distance Classifier

• Concept: Measures the Euclidean distance between an unknown pixel and the mean (centroid) of each training class.

• Decision Rule: The pixel is assigned to the class with the closest mean.

Parallelepiped Classifier

Creates a box-shaped decision boundary using min-max values from training samples for each band. If a pixel falls within the box, it is assigned to that class. If it falls in multiple boxes, it remains unclassified or is assigned based on a rule (e.g., closest mean).

Nearest-Neighbor Classifier

To classify an unknown pixel into m classes, the classifier computes the Euclidean distance of the pixel to be classified to the nearest training data pixel.

It could use a majority rule, “the nearest group of pixels”

Gaussian Maximum Likelihood Classifier (MLC)

Assumes each class follows a Gaussian (normal) distribution and assigns a pixel based on the highest probability (likelihood) of belonging to a class.

Benefits: Takes into account variance and covariance

Downsides: computationally expensive, assumed normal distribution for classes

When to use each supervised classification type

• Use Minimum Distance and Parallelepiped for fast classification but only when class boundaries are well-defined.

• Use Nearest-Neighbor (k-NN) if class distributions are complex and not Gaussian.

• Use Maximum Likelihood (MLC) for the highest accuracy when you have large training samples and Gaussian-distributed data.

Chain method

1. Program reads the data and builds clusters.

2. A minimum distance to mean approach is used to associate each pixel to a cluster

K-means

1) Initial mean (seed) specified for K clusters.

2) Pixels closest to this mean are assign to each cluster

3) Reiteration (migration of pixels)

sources of variability that influence classification accuracy (4)

1. Sensor-Related Variability (low spatial resolution, few bands, low radiometric resolution (bits) → less info = less accurate categories

2. Environmental conditions ie cloud cover, shadows from topographic features

3. Training Data: misclassification, not enough training data, similar spectral responses (building v dry soil)

4. Model Choosing a parametric model for non parametric data