A neural network approach to cloud removal in single-band SSM/I imagery

摘要

When mappign geophysical variables (e.g. ground albedo) with satellite imagery, it is common practice to create composite maps depicting seasonal or monthly temporal averages. Using data from multiple satellite passes, each location on the earth is represented by a vector of several measurements acquired over the compositing period. Removing the cloud-contaminated values from each measurement vector is a common preprecessing task. The objective of this research is to detect and remove hydrometeor contamination in time-composite 85 GHz (vertically polarized) SSM/I data of the Amazon Basin without reference to any other SSM/I channel. To develop the cloud removal algorithm, a feed-forward neural network was trained using 85 GHz SSm/I brightness values produced through simulation. Since the data was synthetic rather than real, the correct mean brightness value of each vector was known, as well as the level of contamination for each measurement in the vector. The network inputs included several measures designed to emphasize the atypical cool measurements diagnostic of hydrometeor contamination. The desired output of the network was a binary flag indicating whether the presented measurement was contaminated or clean. When the network was tested, 92percent of the synthetic measurements were correctly classified as clean or dirty. The average error remaining in the decontaminated vectors was less than 0.1K.