Image classification assigns every pixel (or object) in an image to a category — water, forest, cropland, built-up — turning imagery into land-cover maps. Methods range from classic supervised and unsupervised statistics through object-based image analysis to modern deep learning. Whatever the algorithm, the discipline is constant: good training data in, rigorous accuracy assessment out.
Supervised and unsupervised approaches
Supervised classification learns from analyst-labelled training samples: the algorithm — maximum likelihood historically, random forests and gradient boosting today — learns each class’s spectral signature and applies it image-wide. Its quality is bounded by the training data’s quality and representativeness.
Unsupervised classification clusters pixels by spectral similarity without labels, leaving the analyst to interpret clusters afterwards. It is a fast reconnaissance tool and a check on assumptions, less often the final product.
Objects, time series and deep learning
Object-based image analysis (OBIA) first segments imagery into meaningful regions, then classifies them using spectra plus shape, texture and context — essential at high resolution, where single pixels of a roof or tree mean little alone. Time-series classification adds phenology: crops that look identical in one image separate cleanly by their seasonal curves.
Deep learning — semantic segmentation networks in particular — now leads for extracting buildings, roads and complex classes from high-resolution imagery, at the price of far larger training-data appetites.
Accuracy assessment: the non-negotiable step
A classification without accuracy assessment is an illustration, not a measurement. Standard practice reserves independent reference samples, builds a confusion matrix, and reports overall accuracy plus per-class user’s and producer’s accuracy — revealing exactly which classes are trustworthy and which confuse.
Serious programmes design the sample statistically (stratified, sized for target confidence) and document it, so area estimates carry defensible error bounds — the difference between a map and evidence.
Frequently asked questions
What accuracy is considered good for land-cover classification?
Conventionally 85% overall accuracy is a common benchmark, but per-class accuracy matters more: a map can score 90% overall while its rarest, most decision-critical class is barely better than chance. Always read the confusion matrix.
How much training data does classification need?
Classic machine learning may need a few hundred well-chosen samples per class; deep-learning segmentation typically wants thousands of labelled examples or extensive annotated polygons. Quality and representativeness beat raw quantity in both regimes.
Related topics
Need this done professionally?
GLOBEIR delivers this work for government and enterprise clients.