Remote Sensing - Image Classification

Image Classification

• Assigns pixels to specific classes or themes based on spectral characteristics.
• Groups similar pixels by analyzing their reflectance patterns across different spectral bands.

Point / Spectral Classifiers

• Classify individual pixels based on recognized spectral patterns.
• Limited in identifying relationships distinguishable to the human eye.

Neighbourhood / Spatial Classifiers

• Utilize both spectral and textural information for classification.
• Provide higher accuracy compared to spectral classifiers.

Feature Space

• Reflectances are plotted as x; y scatter plots.
• Distances between classes in feature space are crucial for distinguishing classes.

Informational Classes

• User-defined classes based on specific interests and objectives.
• Characteristics such as brightness, topography, color, and texture are considered.

Spectral Classes

• Comprised of uniform pixels based on brightness across multiple channels.
• Spectral classes contribute to the formation of informational classes.

Unsupervised Classification

• Minimal user interaction; spectral classes are grouped initially.
• Algorithms like K-Means and ISODATA identify natural groupings.
• Analyst defines the number of classes and separation, reducing human error risks.
• Cannot align with informational classes; requires additional processing.

ISODATA

• An iterative self-organizing data analysis technique for unsupervised classification.
• Pixels are grouped around centroids and reclassified based on new centroid positions.
• Iterations can be limited or terminated upon classification stability.

Supervised Classification

• Operator defines classes during a training process, relying on spectral information.
• Each pixel is assigned to the class it most closely resembles.
• Can identify classification errors but may lack representative training areas.

Training Areas

• Representative sample sites used for comparison against classification algorithms.
• Recommend using over 100 pixels and 5-10 training areas per class.
• Need for uniformity and homogeneity in training areas.

Parametric Signature

• Based on statistical parameters of training areas (min, max, mean).
• Decision Rules include Mahalanobis Distance, Minimum Distance to Mean, and Maximum Likelihood.

Nonparametric Signature

• Based on defined Areas of Interest in feature space.
• Pixels assigned based on their position within the feature space using Parallelepiped decision rules.

Parallelepiped

• Simplistic classifier that assigns pixels to defined classes based on training data.
• Unclassified pixels are labeled as "unknown," leading to potential class overlaps.

Minimum Distance to Mean

• Measures Euclidean distances from training data to cluster centers (mean).
• Pixels assigned to the nearest cluster but may overlook variability, risking overlap.

Mahalanobis Distance

• Accounts for variations among samples by applying a co-variance matrix.
• Estimates distances from the mean while assuming normal distribution; more computation-intensive.

Maximum Likelihood

• Utilizes training data to estimate means and variances for probability assignments.
• Pixels classified in line with the highest probability; assumes equal class probabilities.

Hybrid Classification

• Combines outputs from unsupervised to train supervised classification models.
• An approach to leverage advantages of both classification types for improved accuracy.