Remote Sensing: Digital Image Classification and Land Cover Change Detection

Vocabulary flashcards covering foundational concepts, classification algorithms, evaluation plots, and land cover change detection techniques from remote sensing lectures.

Digital Image Classification

The process of assigning individual pixels within a multi-band remote sensing dataset to specific thematic categories or informational classes based on statistical similarities in their spectral signatures.

Supervised Classification

An image classification approach where the analyst knows the identity and location of specific land cover types a priori and selects them as training sites to calibrate algorithms that classify all remaining pixels.

Unsupervised Classification

An automated classification method where land cover classes are not known a priori; computational algorithms cluster pixels (number decided by analyst) based purely on natural groupings of spectral characteristics, after which the analyst re-labels clusters into informational classes a posteriori.

Training Sites

Homogeneous, representative sample areas of known land cover types selected by an analyst using field observations, aerial photos, or known optical properties to extract statistical parameters for supervised classification.

Minimum Distance Classifier

A classification algorithm that calculates the Euclidean distance d_{x,A} = \root \relax \nofile \relax \relax using the formula $\sqrt{\sum_{i=1}^{m} (DL_{x,i} - DL_{A,i})^2}$ between an unknown pixel $x$ and the mean center of each candidate class $A$ across all bands $i$, assigning the pixel to the class with the smallest distance.

Parallelepiped Classification

A decision-rule classifier based on multidimensional rectangular bounding boxes defined in feature space by class training means and dispersion ranges ($DL_{A,i} - R_{A,i} \le DL_{x,i} \le DL_{A,i} + R_{A,i}$), which can result in pixels remaining unclassified or falling into overlapping categories.

Maximum Likelihood Classifier

A probabilistic classification algorithm that computes probability density functions based on training mean and variance, assigning each pixel to the class with the highest statistical likelihood under the assumption of a normal distribution.

Feature Space Plot

A scatter plot that graphs pixel brightness distributions across two or more spectral bands simultaneously to visually and statistically evaluate the spectral separability between candidate land cover classes.

Observational Coherence

The prerequisite set of consistent imaging conditions needed for quantitative change detection, requiring images to have the same spectral bands, viewing geometry, geographic coverage, sky conditions, and temporal window (time of day and season).

Image Differencing

An image algebra technique in which digital counts, surface reflectance, or vegetation indices from two dates are subtracted pixel-by-pixel, allowing analysts to apply histogram thresholds to separate change (loss or gain) from stable pixels.

Normalized Burn Ratio (NBR)

A spectral index defined as $NBR = \frac{NIR - SWIR}{NIR + SWIR}$, used to identify burned areas and evaluate burn severity by calculating differences ($\Delta NBR$) between pre-fire and post-fire imagery.

Normalized Difference Wetness Index (NDWI)

A remote sensing index defined as $NDWI = \frac{\text{Green} - NIR}{\text{Green} + NIR}$, designed to delineate open water features by exploiting strong near-infrared absorption and green reflectance to separate water from soil and terrestrial vegetation.

Multitemporal Color Composite

An image visualization technique where identical spectral bands or indices from two or three distinct dates are assigned to the red, green, and blue (RGB) color channels, highlighting persistent land cover in neutral tones and areas of change in distinct hues.

Multidate Layer Stacking

A change detection procedure where coregistered bands from multiple observation dates are combined into a single unified multiband data stack and categorized directly using supervised or unsupervised classification algorithms.

Post-Classification Comparison

A change detection method where images from multiple dates are classified independently into discrete thematic maps and then compared on a pixel-by-pixel basis, reducing sensitivity to sensor or atmospheric differences but compounding individual map classification errors.

From-To Change Detection Matrix

A cross-tabulation table produced in post-classification comparison that tracks every initial land cover class at Date 1 against every resulting class at Date 2 to identify and map specific land transition trajectories.

Spectral Change Vector Analysis (CVA)

A multidimensional change detection technique that models pixel spectral displacement between two dates as a vector in spectral space, evaluating both the magnitude (intensity of change) and angle (direction or type of transition).

Time Series Analysis (Remote Sensing)

A change detection method using dense, continuous temporal observation sequences (facilitated by the open Landsat archive established in October 2008) to fit algorithmic trajectories and detect discrete change points, gradual recovery, or cyclical phenology.