Cloud Detection Based on Support Vector Machine for Landsat 8 Imagery

Abstract

Cloud covers are generally present in optical remote-sensing images which limit the usage of acquired images and increase the difficulty of data analysis such as image compositing correction of atmosphere effects calculations of vegetation indices land cover classification and land cover change detection Therefore the generation of cloud mask is one of important pre-processing steps in many remote-sensing researches In previous studies thresholding is a common and efficient method in cloud detection However a selected threshold is usually suitable for certain cases or local study areas and it may be failed in other cases In other words thresholding-based methods are data-sensitive Besides there many exceptions have to control and the Earth environment and atmosphere changed dynamically Using a set of threshold values on various image data is not effective In this study a threshold-free method based on Support Vector Machine (SVM) is proposed which can avoid the abovementioned problems A statistical model is adopted to detect clouds instead of a subjective thresholding-based method which is the main idea of this study The features used in a classifier is the key to a successful classification In the popular thresholding-based approach called Automatic Cloud Cover Assessment (ACCA) the cloud is distinguish from other objects based on the physical characteristics of cloud and other targets on the ground Similarly the algorithm called Fmask adopted a lot of thresholds and criteria to screen clouds water and snow Following these two algorithms the spectral features used in the proposed method is defined by the ACCA and Fmask algorithms Spatial information is also important in the classification processing in addition to the spectral information Consequently co-occurrence matrix of the Hotelling transform is used in proposed method to extract the spatial or called texture features In this study images are partitioned into four groups: cloud snow water and the others In experiments images acquired by the Landsat 8 Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS) and images containing the landscapes of agriculture snow area and island are tested Experiment results demonstrate the overall detection accuracy of the proposed method is about 93% to 97% which is better than thresholding-based methods with the default threshold values and is comparable to that with the tuned threshold values

Date of Award	2016 Sept 7
Original language	English
Supervisor	Chao-Hung Lin (Supervisor)