Evaluating and reducing errors in seasonal profiles of AVHRR vegetation indices over a Canadian northern national park using a cloudiness index

High-temporal coarse resolution remote-sensing data have been widely used for monitoring plant phenology and productivity. Residual errors in pre-processed composite data from these sensors can still be substantial due to cloud contamination and aerosol variations, especially over high cloud-cover areas such as the Arctic. Commonly used smoothing and filtering methods try to reform the often heavily distorted seasonal profiles of vegetation indices one way or another, instead of explicitly dealing with the errors that cause the distortion. As the distortion varies from year to year for a pixel or from pixel to pixel, so does the performance of various smoothing and filtering methods. Consequently, change detection results are likely method dependent. In this study, we investigate alternative methods in order to eliminate bias caused by cloud contamination and reduce random errors due to aerosol variations in the 10 day Advanced Very High Resolution Radiometer (AVHRR) composite data, so that accurate seasonal profiles of vegetation indices can be constructed without the need to apply a smoothing and filtering method. The best alternative method corrects cloud contaminations by spatially pairing averages of simple ratio over cloud-contaminated and clear-sky pixels in a class (SPAC). The SPAC method eliminates bias caused by cloud contamination and reduces the relative random errors to <14% near the start/end of a growing season, and to <8% during the middle growing season for the six treeless wetland and tundra classes in Wapusk National Park. In comparison, with the method whereby all pixels in a class (average all pixels in the class (AAC)) are averaged in a period, the bias could be up to 40% if all the pixels in the composite period are heavily cloud contaminated.