Cloud and cloud shadow screening across Queensland, Australia: An automated method for Landsat TM/ETM + time series

Abstract

The 20+ year collection and moderate resolution of Landsat Thematic Mapper (TM) imagery and Enhanced Thematic Mapper (ETM+) imagery provide a crucial data source for analysing land surface change over time for a range of applications. In Queensland, Australia, a number of government policies, natural resource management programmes and research activities are reliant on large-area, multi-temporal land cover monitoring applications based on Landsat satellite imagery. However, clouds and associated cloud shadows frequently obstruct the view of the land surface. The restriction of analyses to cloud-free imagery will reduce the opportunities to sample the land surface and limit the analysis of trends in reflectance over time. This study presents a new automated method to screen cloud and cloud shadow and is intended for application to entire time series of Landsat imagery rather than single images processed in near- real time. The method uses a hierarchical approach and takes advantage of spectral, temporal, and contextual information. Outliers are located relative to the time series of land surface reflectance by smoothing time series information using minimum and median filters which are then used in multi-temporal image differencing. Seeded region grow and morphological dilation (pixel buffering) filters are then applied to map a larger spatial extent of the cloud/cloud shadow. Spectral and contextual rules were developed empirically using calibration and validation data derived from six Landsat WRS Path/Rows (number of images=60) with varying climatic and land surface characteristics across the state of Queensland, Australia. The validation demonstrates that cloud contaminated pixels were accurately classified with producer's, user's and overall accuracies of 98, 87 and 97%, respectively. The ability to detect cloud shadow was less accurate, in comparison, with producer's, user's and overall accuracies of 90, 62 and 97%, respectively. The pixel buffer was found to be the largest source of commission error for cloud and cloud shadow in the final classifications. However, for many applications removing additional cloud/shadow at the expense of higher commission errors may be desirable. The performance of the method was also compared with the published Fmask (Function of mask) method. This demonstrated a moderate improvement in the detection of cloud (producer's accuracies: time series 98% and Fmask 90%; and an equivalent user's accuracy of 87%), and a significant improvement in the detection of cloud shadow (producer's accuracies: time series 90% and Fmask 78%; user's accuracies: time series 62% and Fmask 50%). Importantly, the results indicate that this automated method is robust and that temporal information can improve the detection of cloud and cloud shadow, although shadow detection above cropping areas is limited. The calibration/validation of the method has been restricted to Queensland, Australia. With further development there is potential for this method or one using a similar framework to have wider application in other landscapes.