A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Image classification assigns each pixel (or object) in a remote sensing image to a land cover or land use category based on its spectral, spatial, or temporal characteristics. Supervised classification trains an algorithm on labeled samples (training data) where the analyst has identified known examples of each class; the algorithm then extends these labels to the entire image. Unsupervised classification groups pixels by spectral similarity without prior labels, and the analyst interprets the groups afterward. Modern approaches include object-based classification (grouping pixels into meaningful segments first), deep learning (convolutional neural networks), and multi-temporal classification using time-series phenology.
The fundamental goal of most remote sensing projects is not just to look at pretty images but to convert imagery into thematic information -- maps showing what is on the ground. Image classification is the core technique for this conversion, transforming continuous spectral data into discrete categories like forest, cropland, water, or urban.
Supervised classification follows a workflow: collect training samples (pixels with known labels), extract their spectral signatures, train a classifier (maximum likelihood, random forest, support vector machine, or neural network), apply the classifier to the full image, and validate results with independent test data. The quality of training data largely determines classification accuracy -- the algorithm cannot learn distinctions the training data does not represent.
Unsupervised classification takes the opposite approach. Algorithms like K-means or ISODATA cluster pixels into groups based on spectral similarity alone, without any labeled data. The analyst then examines each cluster and assigns it a thematic label. This is particularly useful for initial exploration of unfamiliar imagery, discovering spectral classes that may not correspond to predefined categories.
The accuracy assessment is as important as the classification itself. A confusion matrix compares classified labels against reference data, yielding overall accuracy, producer's accuracy (how well each class is detected), user's accuracy (how reliable each class label is), and kappa coefficient (accounting for chance agreement). Without rigorous accuracy assessment, a classification map has no quantified reliability and cannot support decision-making.