A KNOWLEDGE-BASED APPROACH FOR REDUCING CLOUD AND SHADOW

Abstract

Completely cloud-free remotely sensed images are not always available, especially in tropical, neo-tropical, or humid climates, posing complications and perhaps serious constraints to image analysis. In this paper, a knowledge-based approach to reduce cloud and shadow using two dates of Landsat TM imagery is presented. Procedurally, first, the brightness and contrast of a secondary image was adjusted to be the same as the main image. Second, instead of detecting clouds and shadows in two images independently, a knowledge base was used to detect their presence in the main image in areas not present in the secondary image. Corresponding brightness value thresholds for cloud and shadow area in the main image and non-cloud and non-shadow area in the secondary image were set in TM bands 1 and 4, respectively. Thresholds of brightness difference between two images were used to avoid other features having similar brightness values. An object-based method was employed to distinguish linear water features from shadow areas. eCognition was used to segment images into objects, with emphasis on water and shadow features. and length-to-width ratio of these objects was derived to separate confused linear water and shadow area. Finally, a composite image was generated with minimal cloud and shadow by replacing the brightness values of detected areas in the main image with those of the secondary image. This paper describes the method and presents results for a study area in Madagascar.