Key issues in image understanding in remote sensing

Abstract

Remotely sensed images of a planet’s atmosphere, oceans and surface contain a plethora of confusing signals about the physical nature of these phase states. Historically, there has been an emphasis on the semi-automated extraction of feature classes based on the spectral properties of objects viewed within a scene and on the use of ad hoc manual photointerpretation techniques. Although these approaches will remain important, they are inadequate, on grounds of speed, accuracy and cost, for the increasing demands of data-gatherers and consumers. Research has recently begun into the automation of image-interpretation tasks and the development of parallel machines with the required processing capabilities. Three important requirements are: (i) means to simulate the appearance of a scene, including the interaction of electromagnetic radiation with the surface and the effects of any intervening atmosphere; (ii) an understanding of how knowledge can be captured and introduced at different levels in the processing hierarchy and (iii) the application of constraints based on a knowledge of the geometry of objects in the scene. These three aspects will be illustrated by examples from various fields, including petroleum exploration, measurement of fluid motion and the extraction of digital terrain elevation models.