5.17.3.A. Supervised Classification

Supervised classification consists of three distinct stages; training, allocation and testing, as shown in Figure 5.44.

Supervised classification is a method used to classify image data based on known categories. This classification process is divided into three stages: training, allocation, and testing. The training stage involves selecting sample pixels from the image that represent known classes of land cover, such as residential or agricultural areas. In the allocation stage, software uses the data from the training stage to classify the remaining pixels in the image based on their spectral signatures. Finally, testing involves evaluating the accuracy of the classification against actual reference data.

Imagine you are a teacher preparing a standardized test for your students. You first go through the lessons (training) to ensure you understand the material fully before creating questions (allocation) for a test. After giving the test, you check the results against the students' actual understanding (testing) to see how accurately the test assessed their knowledge.

Training is the first stage where the identification of a sample of pixels of known classes is done with the help of reference data, such as field visits, existing maps and aerial photographs.

In the training phase, the focus is on identifying sample pixels from the image that represent specific, known categories. These pixels are selected using reference data like existing maps or direct field observations. The digital numbers (DN) associated with these pixels help define their characteristics. This step is crucial because the software later uses this information to learn how to classify other pixels in the image accurately.

Think of this training phase as teaching a child to identify animals. You might show them pictures of a dog, a cat, and a cow and explain the features of each animal. By doing this, the child learns what makes each animal unique. Later, when shown a new picture, they will use that prior knowledge to identify the animal correctly.

In the second stage, the training pixels are used by the software to derive various statistics for each class, and are correspondingly assigned signatures.

Once training pixels are identified, the software analyzes their DN values to create statistical characteristics for each class. This involves examining the variability in DN values for each training area to develop a 'signature' or a profile that describes how that land cover type looks in the image. This signature serves as a reference for classifying similar pixels throughout the rest of the image.

This process is similar to a student learning how to recognize different musical notes. When a music teacher plays a note, the student memorizes its tone (the signature). Later, when they hear that note again in a song, they can identify it just by recognizing its unique sound compared to others.

In the third stage, the remaining pixels of the image are allocated to the same class with which they show greatest similarity based on the established signature files in the second stage.

After training and defining statistical signatures, the software then classifies each unclassified pixel in the image by comparing it to the established signatures from the training phase. This allocation process identifies which class a pixel most closely resembles based on its DN characteristics. The end result is a classified image where every pixel is assigned to its corresponding land use or cover class.

Think of a chef preparing a large meal. They've tasted several dishes (training) and know the flavors. As they cook, they taste each new dish (remaining pixels) and determine which recipe it most closely resembles (allocation). By the end, every dish is categorized and ready to be served.

It is important to note that training sites developed in one scene may or may not be replicable to entire study area due to variation in ground objects and conditions.

The training phase can influence the overall classification results significantly. If the selected training sites do not represent the entire area accurately due to variations in conditions like illumination or atmospheric effects, the classification may fail to produce reliable results. Therefore, iterative refinement is often necessary, where the analyst continues to adjust the training samples until satisfactory accuracy is achieved.

This situation is akin to a photographer trying to capture the perfect landscape shot. If they only practice in one type of light or weather, their skills may not translate to another scenario. To ensure they can adjust for different environments, they must practice in various conditions to become versatile and adapt their techniques, just like analysts must refine training sites.

--