SUMMARYWhen using composite optical satellite images for land studies, an accurate and sensitive method is needed to detect pixels contaminated by unwanted atmospheric and surface effects. In this paper, we have examined the feasibility of using an algorithm previously developed for post-season analysis but in a forward' mode, i.e. for identifying contaminated pixels in current-season data. The CECANT algorithm (Cloud Elimination from Composites using Albedo and NDVI Trend; Cihlar, 1996) uses AVHRR channel 1 reflectance to detect strongly contaminated pixels (bright clouds, snow), and the normalized difference vegetation index (NDVI) to identify partially contaminated pixels over land. However, this approach needs a complete NDVI seasonal trajectory to derive the adaptive thresholds. Since some important applications imply near-real time processing, we have examined the possibility of deriving the thresholds from previous (historical) AVHRR data. Using four years of AVHRR data of Canada we found the accuracy to vary (55–100%), with thresholds from a single year yielding anomalous results in some compositing periods. On the other hand, the use of thresholds derived from averaged data (over 3 years in this case) produced more accurate and consistent results. Assuming that contamination masks derived from each year's data are 100% accurate, the masks based on the three years of AVHRR data were 75–95% correct (4-year average), depending on the compositing period. Two alternative adjustments were attempted to improve the accuracy and consistency of the identification of contaminated pixels. By relaxing one of the CECANT thresholds, the errors of omission and commission could be more equally balanced (∼10 and 20%, respectively). Further improvement was obtained by adjusting the NDVI data to the data set initially used in deriving the threshold coefficients, and then deriving a new set of coefficients for that period. In this case, the omission and commission errors were nearly equal and, importantly, the performance was consistent among all the years. The performance of an alternative, simplified cloud screening method (based on fixed reflectance and temperature thresholds) was also examined for a comparison, and was found to have lower accuracy (58% compared to 87%, 4-year average). It is concluded that CECANT may be effectively applied in near-real time processing of satellite optical data. However, the residual mismatches indicate that a reprocessing after the end of the growing season may be desirable for applications requiring high radiometric accuracy.