Glossary of Key Terms

Sensor and Data Concepts

• Image Elements/Pixels: Digital representation of sampled object surfaces through the Instantaneous Field of View (IFOV) of sensor arrays.
• Digital Numbers (DNs): Photon counts recorded by sensor arrays in distinct spectral bands, which can be converted into spectral features. Range from 0 to 255 for 8-bit data.
• Ground Resolution Cell (GRC): The area on the ground represented by a single pixel.
• IFOV: The angle of the solid cone from the sensor that intersects an object surface, projecting onto a sensor element. Not constant; it depends on Y (the height of the sensor element) and F (the focal length).

Photon Capture and Digital Numbers

• Photon Stream/Flux: The number of photons falling onto or into a sensor element during an exposure time.
• Electro-Optical Devices: Imaging sensors are "arrays of cells that convert photons into electronic signals".
• Digital Numbers (DNs): The sensor counts photons, and these counts are "Digital Numbers" in distinct spectral bands. This photon count is converted to an analog voltage (from accumulated charge in a capacitor), which is then converted to digital data via an Analog-to-Digital Converter (ADC). For an 8-bit record, DNs range from 0 to 255.
• Measurement Vector: For a given image sample position, the set of N measurements from different spectral bands forms an N-dimensional "measurement vector" d = [d1 d2 … dN]T.

Radiometric Properties and Classes

• Radiometric Measurements: The primary outputs of a sensor, such as an RGB camera or a multispectral scanner like Sentinel-2.
• Radiometric Classes: A finite set of possible class labels (Ω = {ω1, ω2,…, ωK}) that pixels are assigned to. These are defined by the GIS user community (e.g., {Grass, Wheat, Potatoes, Sugar-beets, Beans, Peas, Onions} for agricultural classification). An "unknown class" (ω0) encompasses features not included in the defined classes.
• Box Classifier (Level-Slice Classifier/Parallelepiped Classifier): The simplest classification method that defines upper and lower limits for each band and class, creating box-like areas in feature space. Pixels falling within a box are assigned its class.
• Cluster: A group of adjacent points in feature space that are spectrally similar, representing a distinct radiometric or land cover class.
• Error Matrix (Confusion Matrix/Contingency Matrix): A table that compares the results of an image classification to reference (ground truth) data, characterised classification accuracy.
• Feature Space: A multi-dimensional graph or scatter plot where the Digital Numbers (DNs) from multiple spectral bands for each Ground Resolution Cell (GRC) are plotted as a single point (feature vector).
• Feature Vector: A vector composed of the Digital Numbers (DNs) from different spectral bands for a single Ground Resolution Cell (GRC).
• Ground Resolution Cell (GRC): The area on the ground represented by a single pixel in a remote sensing image.
• Hard Classification: An image classification method that assigns each pixel exclusively to one single class.
• Image Space: The 2D array of pixels that constitutes a digital image, where the spatial distribution of Digital Numbers (DNs) defines the image.
• Instantaneous Field of View (IFOV): The angle or solid cone from a remote sensing sensor that projects onto a single sensor element, determining the area on the ground represented by a pixel.
• Iterative Optimisation (Migrating Means/ISODATA): An unsupervised clustering algorithm that iteratively adjusts cluster centres (means) and reassigns pixels until cluster centres stabilise.
• Kappa (κ) Coefficient: A statistical measure of classification accuracy that takes into account agreement occurring by chance, providing a more robust assessment than simple overall accuracy.
• Knowledge: In remote sensing, based on hypotheses and measured evidence, often expressed as the probability of a hypothesis given the evidence.
• Maximum Likelihood (ML) Classifier: A commonly used classification algorithm that assigns a pixel to the class for which it has the highest probability, considering not only cluster means but also the shape, size, and orientation of clusters based on their statistical distributions.
• Measurement: The interaction between two objects in space-time that produces data about that interaction, such as a photon sensor measuring photon counts.
• Multispectral Data: Remote sensing data collected across multiple distinct spectral bands of the electromagnetic spectrum.
• Overall Accuracy: The total number of correctly classified pixels divided by the total number of pixels in the validation sample, typically derived from an error matrix.
• Pattern: The spatial arrangement of objects and the characteristic repetition of certain forms or relationships in an image.
• Photon Flux: The number of photons falling on or into a sensor element per unit time.
• Photons: Fundamental particles of light, carrying electromagnetic energy.
• Pixel (Image Element): The smallest unit of a digital image, representing a Ground Resolution Cell (GRC) and containing Digital Number (DN) values for each spectral band.
• Pixel-based Classification: An image classification method that processes and assigns each individual pixel to a class based primarily on its spectral information.
• Radiometric Class: A distinct category or label assigned to pixels during image classification, based on their spectral characteristics.
• Region of Interest (ROI): Sample areas selected on homogeneous regions of an image during supervised classification, used to collect training samples for defining class spectral characteristics.
• Remote Sensing (RS) Image: An image acquired by sensors on platforms (e.g., satellites, aircraft) that record electromagnetic radiation reflected or emitted from the Earth's surface.
• Soft Classification: An image classification method that assigns multiple classes to each pixel, each with an associated likelihood or probability, to better account for within- and between-class variation and mixed pixels.
• Spectral Characteristics: The unique ways in which different materials interact with electromagnetic radiation across various wavelengths (e.g., reflectance, absorption).
• Spectral Class: A class defined purely by its spectral characteristics as identified in the feature space, which may or may not directly correspond to a desired land cover type.
• Spectroscopic Data: Data that measures the intensity of electromagnetic radiation at different wavelengths, providing detailed spectral signatures of materials.
• Supervised Classification: An image classification approach where the operator defines the spectral characteristics of classes by identifying sample areas (training areas) in the image.
• Training Areas: Sample areas selected by an operator in supervised classification, representing known land cover types, used to train the classification algorithm.
• True Class: The actual class or category of a feature on the ground, used as a reference for validating image classification results.
• Unsupervised Classification: An image classification approach where clustering algorithms automatically find spectral groupings (clusters) in the data without prior knowledge or training samples from the operator.
• Validation: The process of assessing the quality and reliability of image classification results by comparing them against independent reference data.

Definition of clusters in feature space: Achieved through supervised or unsupervised classification.

Selection of the classification algorithm: Deciding how pixels are assigned to classes.

Running the actual classification: Assigning each multi-band pixel to a predefined class.

Validation of the result: Assessing classification quality against reference data.

Band Selection: Avoiding highly correlated bands (redundant information) and considering hardware/software limitations.