A Contextual Classification Approach for Remote Sensing Image Classification of Hyperspatial Imagery

One of the most important tasks of remote sensing is the use of imagery to classify and delineate different objects on the earth's surface. Conventional methods of pixel-based classification classify each pixel independently by only considering its spectral properties. These pixel-based techniques are most applicable to medium and coarse-scale remote sensing, but often become less accurate at high spatial resolutions (pixels <= 1m) as the scene objects become larger than a pixel. Contextual classification techniques use not only the spectral properties of the pixel, but also the local spatial information to improve pixel labeling and classification. In this study, we use a focal pixel as predictor variables to use with a machine learning classifier. We applied this technique to a set of remotely sensed, multispectral hyperspatial imagery (Worldview-2) to map the type, distribution, and structure of vegetation in a Sierra Nevada forest.