Q&A

Describe the role of "photon counts" in remote sensing data acquisition. How are these physical measurements eventually represented as "Digital Numbers - DN" in an image? Photon counts represent the number of photons captured by a sensor element during an exposure time, indicating the amount of electromagnetic radiation reflected or emitted from a ground resolution cell. These photon counts generate an accumulated charge in the sensor element, which is converted into an analog voltage. This analog voltage is then transformed into a discrete "Digital Number" (DN) by an Analog-to-Digital Converter (ADC) for digital image representation.

Define "image space" and "feature space" in the context of image classification. How do these concepts relate to the photon counts or Digital Numbers (DNs) recorded by an RGB or multispectral sensor? Image space refers to the 2D array of pixels where the spatial distribution of DNs defines the image. Feature space is a multi-dimensional graph where the DNs from different spectral bands for a single Ground Resolution Cell (GRC) are plotted as a "feature vector." Essentially, image space shows the spatial arrangement of DNs, while feature space shows the spectral relationships among them.

Given the challenge of visualising multi-dimensional feature vectors in remote sensing when dealing with a high number of spectral bands, what is a practical solution for displaying the distribution of pixel values? The 2D scatterplots are often used as a practical solution for visualising data when dealing with four or more bands, by plotting all possible combinations of two bands separately.

What is digital image classification in remote sensing, and what are its primary applications?

Digital image classification is a process of assigning pixels in an image to specific classes based on their spectral characteristics. It essentially creates a map where each pixel represents a particular land cover type (class). This is done by analysing the photon counts or DNs of pixels, which represent the electromagnetic radiation reflected or emitted from object surfaces or ground resolution cells (GRCs) across different spectral bands.

Its primary applications are diverse. For example, it's used in projects to verify farmers' claims for crop subsidies, often followed by field checks to ensure accuracy. The goal is either to classify scene objects (e.g., crops, water bodies, urban areas) or to estimate parameters related to these objects.

Briefly describe the core principle behind the Minimum Distance to Mean (MDM) classification algorithm. What is one notable disadvantage of this method, and how can it be mitigated? The MDM classifier assigns a pixel to the class whose cluster centre (mean feature vector) is spectrally closest to the pixel's feature vector. A disadvantage is that it does not account for the variability or shape of the clusters, potentially assigning points far from a cluster's core to that class. This can be mitigated by defining a threshold distance, beyond which a pixel will not be assigned to any class.

What are the differences between supervised and unsupervised image classification methods?

Supervised Image Classification:

Requires prior knowledge: The operator must be familiar with the area of interest and know where to find the specific classes (e.g., "water," "forest," "urban") within the scene. This knowledge often comes from field observations or existing maps.

Training areas: The operator manually selects "training areas" (Regions of Interest or ROIs) on the image that are representative of each desired class. These ROIs define the spectral characteristics and boundaries of the clusters in the feature space.

Operator-driven: The quality of the classification heavily depends on the operator's skill in selecting accurate and representative training samples.

Direct class mapping: The output directly maps to predefined, user-named classes.

Unsupervised Image Classification:

Does not require prior knowledge (initially): Useful when detailed knowledge of the area or specific classes is insufficient.

Automated clustering: Clustering algorithms automatically identify "natural" groupings or clusters of spectrally similar pixels in the feature space. The user typically specifies the maximum number of clusters.

Iterative process: Algorithms like the iterative optimisation (migrating means) technique (also known as ISODATA) start with arbitrary cluster centres and iteratively refine them by assigning pixels to the nearest cluster and recalculating cluster means until stability is achieved.

Post-classification labelling: The resulting clusters are initially statistical groupings. The user must then interpret and label these clusters (e.g., "Cluster 1 is water," "Cluster 2 is vegetation") based on external information or visual inspection.

Spectral groupings: Primarily produces spectral classes, which may then need to be aggregated into meaningful land cover types (classes).

In essence, supervised classification is "training by example," while unsupervised classification is "clustering by similarity."

Explain the concept of "feature vector" and "clusters" in the context of image classification.

A feature vector is a multi-dimensional representation of a single pixel's spectral characteristics. For a multispectral image with 'N' spectral bands, each pixel has 'N' Digital Numbers (DNs), one for each band. This set of 'N' DNs forms a feature vector (e.g., [DN_band1, DN_band2, ..., DN_bandN]). This vector can be plotted as a single point in an N-dimensional "feature space."

Clusters are groupings of these feature vectors (points) in the feature space. The fundamental assumption in image classification is that pixels belonging to the same class (e.g., water, forest, bare soil) will have similar spectral properties. When plotted in feature space, these spectrally similar pixels will naturally group together, forming distinct "clusters." For example, all pixels representing water might form a compact cluster in a specific region of the feature space, while pixels representing bare soil form another, often separate, cluster. The goal of classification algorithms is to identify these clusters and then assign every pixel in the image to the class associated with the cluster its feature vector falls into or is closest to.

How is the quality of image classification results validated and measured?

The quality of image classification results is crucial and is typically assessed through a rigorous validation process, often referred to as accuracy assessment. This involves:

Reference Data (Ground Truth): Comparing the classified image with "ground truth" data. This reference data is preferably derived from direct field observations or from higher-accuracy sources like drone images.

Sampling Scheme: Selecting a statistically representative sample of pixels from the classified image for comparison. Common sampling strategies include simple random sampling or stratified random sampling, considering factors like the number of samples and the size of the sample unit (point or area).

Error Matrix (Confusion Matrix): Once samples are collected and compared, an error matrix is constructed. This table cross-tabulates the classified classes against the reference (true) classes. For example, if there are classes A, B, C, D, the matrix shows how many pixels truly belonging to A were classified as A, B, C, or D, and so on.

Accuracy Measures: Various accuracy measures are derived from the error matrix:

Overall Accuracy: The total percentage of correctly classified pixels.

These measures provide quantitative insights into the reliability and quality of the classification output.