Image Classification

Keywords

artificial neural networks, decision trees, machine learning, maximum likelihood, object oriented classification, support vector machines, spectral mixture analysis.

Introduction

Humans have been able to visually identify different types of land-use and land-cover in aerial photography ever since they were first acquired in the mid-1800s. Human visual image classification is based on the use of the fundamental elements of image interpretation such as size, shape, shadow, tone, color, texture, pattern, site, association, etc. Visual image classification is still very important and is performed by millions of people every day as they browse remote sensing images on the internet (e.g., using Google Earth or Virtual Earth) to identify features of interest.

Since the mid-1960s, humans have been able to extract land-use/land-cover and biophysical information directly from remote sensor data using digital computers and special-purpose image classification algorithms. Digital image classification algorithms may be based on parametric statistics (which assume normally distributed data), nonparametric statistics (which do not require normally distributed data), and non-metric [those that can operate on both real-valued data (e.g., spectral reflectance data with values from 0 to 100%) and nominal scaled data (e.g., category 1 =≤5% slope, category 2 = 6−10% slope)]. Digital image classification algorithms make use of supervised or unsupervised classification logic. Furthermore, the classification algorithms can process individual pixels (referred to as per-pixel classifiers) or groups of contiguous pixels (referred to as object-oriented classifiers). This chapter introduces some of the fundamental types of supervised and unsupervised classification algorithms based on per-pixel and object-oriented classification logic.

Land-use and/or land-cover information derived from remote sensor data should be categorized using a documented classification system such as the U.S. Geological Survey's Land Use and Land Cover Classification System for Use with Remote Sensor Data or the American Planning Association's Land Based Classification System so that the information extracted may be shared. Jensen (2005) summarizes several of the most important image classification schemes.

Supervised Classification

The objective of a supervised classification is to categorize every image pixel into one of several predefined land type classes. The five phases of supervised classification are class definition, preprocessing, training, automated pixel assignment, and accuracy assessment. These phases are seldom performed in isolation. For example, if the accuracy assessment phase reveals that certain landtype classes are poorly distinguished, the pixel assignment phase might be revisited using a different classifier.

The problem of supervised landtype classification is one of pattern recognition. Pattern recognition is sometimes called machine learning. A classifier is an algorithm that learns patterns associated with a set of landtype categories and then uses that knowledge to place new patterns into one of the learned categories. The pattern or pattern vector (x) is a set of characteristics used to classify the pixel. The individual measurements in the pattern vector are called features. The typical feature pattern for a landtype category is often referred to as its signature. Features typically include reflectance values from the several image bands (e.g., green, red, and near-infrared reflectance) or new transformed spectral information created based on principal component scores, vegetation index values or texture measures. Sometimes features include ancillary data such as pixel slope and elevation measured from a digital terrain models.

...