Skip to main content

Supervised Classification

Image Classification in Remote Sensing

Image classification in remote sensing involves categorizing pixels in an image into thematic classes to produce a map. This process is essential for land use and land cover mapping, environmental studies, and resource management. The two primary methods for classification are Supervised and Unsupervised Classification. Here's a breakdown of these methods and the key stages of image classification.


1. Types of Classification

Supervised Classification

In supervised classification, the analyst manually defines classes of interest (known as information classes), such as "water," "urban," or "vegetation," and identifies training areas—sections of the image that are representative of these classes. Using these training areas, the algorithm learns the spectral characteristics of each class and applies them to classify the entire image.

  • When to Use Supervised Classification:   - You have prior knowledge about the classes.   - You can verify the training areas with ground truth data.   - You can identify distinct, homogeneous regions for each class.

Unsupervised Classification

Unsupervised classification, on the other hand, uses the spectral properties of the image data to automatically group pixels with similar spectral characteristics into spectral classes. These classes are later labeled by the analyst based on the spectral patterns and ground-truth information.

  • When to Use Unsupervised Classification:   - You have limited prior knowledge about the image's content.   - You need a large number of classes or wish to explore the data's spectral characteristics.   - It's beneficial for quickly exploring unknown regions.

2. Key Stages of Image Classification

Image classification follows a systematic series of stages to produce accurate thematic maps.

  1. Raw Data Collection: Initial, unprocessed image data is collected.
  2. Preprocessing: Prepares the data for analysis by correcting atmospheric effects, removing noise, and aligning geometry. This stage is essential to ensure data accuracy.
  3. Signature Collection: In supervised classification, the analyst collects samples, called signatures, representing each class. These signatures capture the typical spectral characteristics for each category.
  4. Signature Evaluation: The quality and distinctiveness of signatures are evaluated to ensure that they are statistically separate and represent the classes accurately.
  5. Classification: Using the collected signatures, the classification algorithm assigns each pixel to a specific class, producing the classified map.

3. Information Class vs. Spectral Class

  • Information Class: An information class represents real-world categories, such as water bodies, urban areas, or vegetation, specified by the analyst for extraction from the image.
  • Spectral Class: A spectral class is determined by the clustering of pixels with similar spectral (color or brightness) values. These classes are automatically identified based on statistical similarities in pixel values across multiple spectral bands.

4. Supervised vs. Unsupervised Training

To classify an image, a system needs to be trained to recognize patterns.

  • Supervised Training:   - Controlled by the analyst, who selects representative pixels and instructs the system on what each class should look like.   - Often more accurate but requires skill and understanding of the region.
  • Unsupervised Training:   - The computer automatically groups pixels based on spectral properties, with the analyst specifying the desired number of classes.   - This approach requires less skill but may be less accurate.

5. Classification Decision Rules in Supervised Classification

In supervised classification, different decision rules guide the process of assigning pixels to classes. Here are some common ones:

Parametric Decision Rules

These rules assume that pixel values follow a normal distribution, which allows the system to use statistical measures for classification.

  • Minimum Distance Classifier:   - Calculates the distance between a candidate pixel and the mean of each class signature.   - Assigns the pixel to the class with the shortest distance (e.g., Euclidean or Mahalanobis distance).
  • Maximum Likelihood Classifier:   - Considers both variance and covariance within class signatures.   - Assumes a normal distribution and assigns pixels to the class with the highest probability of belonging.

Nonparametric Decision Rules

These rules do not assume a specific distribution.

  • Parallelepiped Classifier:   - Uses minimum and maximum values for each class and assigns pixels within these limits to the corresponding class.

  • Feature Space Classifier:   - Analyzes classes based on polygons within a feature space, which is often more accurate than the parallelepiped method.


Summary Table

AspectSupervised ClassificationUnsupervised Classification
DefinitionUses predefined classes and training areas.Uses statistical groupings based on spectral properties.
ClassesInformation Classes: Known classes defined by the analyst.Spectral Classes: Classes identified by the system.
Training ProcessAnalyst selects and verifies classes.System automatically groups pixels; analyst labels classes.
Best Use CaseWhen classes are known, distinct, and verifiable with ground truth.When classes are unknown or when exploratory analysis is needed.
Accuracy and Skill RequirementHigh accuracy; requires skill and knowledge.Generally lower accuracy; requires less skill.
Decision RulesMinimum Distance, Maximum Likelihood, Parallelepiped, Feature Space.Classes grouped by spectral similarity.

https://geogisgeo.blogspot.com/2023/01/minimum-distance-gaussian-maximum.html



PG and Research Department of Geography,
Government College Chittur, Palakkad
https://g.page/vineeshvc

Comments

Popular posts from this blog

Quantitative Revolution in Geography

The Quantitative Revolution was a major change in Geography during the 1950s and 1960s. It introduced mathematical, statistical and scientific methods to understand geographical phenomena. Before this revolution, geographers mainly described places, regions, landscapes and differences between areas. After the revolution, they increasingly used numerical data, statistical analysis, mathematical models and theories to explain why geographical patterns occur and how they may change. For example: Traditional approach: Describe the distribution of population in different regions. Quantitative approach: Use statistical methods to identify population patterns, explain the factors influencing population distribution and predict future population growth. The revolution helped establish Geography as a spatial science, focusing on the location, distribution, relati...

Satellite Remote Sensing: Concepts and Imaging Systems

  Satellite remote sensing relies on detectors (sensors) that measure reflected/emitted electromagnetic radiation from the Earth. How sensors collect data depends on: Spectral coverage → Multispectral (few bands), Hyperspectral (hundreds), Thermal, Microwave. Detector type → Discrete detectors, linear arrays, or area arrays. Scanning mechanism → Scanning mirrors (whiskbroom) vs. linear arrays (pushbroom). Multispectral Imaging Using Discrete Detectors and Scanning Mirrors (Whiskbroom scanners) Principle : A single detector (or a few detectors) measures radiation one pixel at a time. A scanning mirror sweeps across-track (perpendicular to the satellite path) to build up the image line by line. The forward motion of the satellite provides the along-track dimension. Known as a whiskbroom scanner . Characteristics : Good calibration stability. Narrow instantaneous field of view (IFOV). Susceptible to mechanical wear (moving ...

Geometric Correction

When satellite or aerial images are captured, they often contain distortions (errors in shape, scale, or position) caused by many factors — like Earth's curvature, satellite motion, terrain height (relief), or the Earth's rotation . These distortions make the image not properly aligned with real-world coordinates (latitude and longitude). 👉 Geometric correction is the process of removing these distortions so that every pixel in the image correctly represents its location on the Earth's surface. After geometric correction, the image becomes geographically referenced and can be used with maps and GIS data. Types  1. Systematic Correction Systematic errors are predictable and can be modeled mathematically. They occur due to the geometry and movement of the satellite sensor or the Earth. Common systematic distortions: Scan skew – due to the motion of the sensor as it scans the Earth. Mirror velocity variation – scanning mirror moves at a va...

RADIOMETRIC CORRECTION

  Radiometric correction is the process of removing sensor and environmental errors from satellite images so that the measured brightness values (Digital Numbers or DNs) truly represent the Earth's surface reflectance or radiance. In other words, it corrects for sensor defects, illumination differences, and atmospheric effects. 1. Detector Response Calibration Satellite sensors use multiple detectors to scan the Earth's surface. Sometimes, each detector responds slightly differently, causing distortions in the image. Calibration adjusts all detectors to respond uniformly. This includes: (a) De-Striping Problem: Sometimes images show light and dark vertical or horizontal stripes (banding). Caused by one or more detectors drifting away from their normal calibration — they record higher or lower values than others. Common in early Landsat MSS data. Effect: Every few lines (e.g., every 6th line) appear consistently brighter or darker. Soluti...

Spectral Signature vs. Spectral Reflectance Curve

Spectral Signature  A spectral signature is the unique pattern in which an object: absorbs energy reflects energy emits energy across different wavelengths of the electromagnetic spectrum. ✔ Key Points Every natural and man-made object on Earth interacts with sunlight differently. These interactions produce a distinct pattern , just like a "fingerprint". Sensors on satellites record these patterns as digital numbers (DN values) . These patterns help to identify and differentiate objects such as vegetation, soil, water, snow, buildings, minerals, etc. ✔ Examples of Spectral Signatures Healthy vegetation → High reflectance in NIR , strong absorption in red Water → Strong absorption in NIR and SWIR , low reflectance Dry soil → Gradual increase in reflectance from visible to NIR Snow → High reflectance in visible , low in SWIR ✔ Why Spectral Signature Matters It allows: Land cover classification Chan...