论文部分内容阅读
Optically thick clouds preclude optical and thermal wavelength remote sensing of the land surface.Cloud detection algorithms have been developed for systematic application to satellite data but the resulting cloud mask products are often unreliable due to the considerable complexity and variability in cloud types and surface backgrounds.The potential of Landsat data processing to provide systematic continental scale time series products has been demonstrated by the NASA Web-enabled Landsat Data (WELD) project (http://weld.cr.usgs.gov/).Recognizing that cloud detection errors,both of omission and commission will occur,two spatially explicit cloud masks are implemented in the current WELD products: the heritage automatic cloud cover assessment algorithm (ACCA) and a decision tree cloud mask algorithm that generally performs better than ACCA for Landsat ETM+ data.These two cloud detection algorithms,and the majority of other published algorithms,apply reflective and thermal wavelength tests to individual satellite acquisitions i.e.,they are mono-temporal.In this paper we present a multi-temporal cloud detection approach that we show provides an opportunity for improved Landsat cloud detection when applied to weekly WELD 30m products.A key requirement is to be able to differentiate clouds from surface changes in a way that is insensitive to intra- and inter-annual variations.A spatially explicit monthly climatology of NDVI,Haze Optimized Transformation (HOT),and brightness temperature (BT) summary statistics are derived from 7 years of weekly WELD product data over the United States.In this way,for each 30m pixel 12 monthly median and median absolute deviation (MAD) NDVI,BT and HOT values are generated.The cloud detection is implemented for each pixel using a nonparametric statistical measure to determine if the difference between the observed weekly WELD product NDVI,BT and HOT values and the monthly climatological equivalents denote a significant change indicative of a cloud.Results are shown across the United States and the statistical comparisons with the current WELD cloud algorithms are presented.