The Effects of Cloud Inhomogeneities Upon Radiative Fluxes, and the Supply of a Cloud Truth Validation Dataset

摘要

The ASTER polar cloud mask algorithm is currently under development. Several classification techniques have been developed and implemented. The merits and accuracy of each have been examined. The classification techniques investigated include fuzzy logic, hierarchical neural network, and a pairwise histogram comparison scheme based on sample histograms called the Paired Histogram Method. Scene adaptive methods also were investigated as a means to improve classifier performance. The feature, arctan of Band 4 and Band 5, and the Band 2 vs. Band 4 feature space are key to separating frozen water (e.g., ice/snow, slush/wet ice, etc.) from cloud over frozen water, and land from cloud over land, respectively. A total of 82 Landsat TM circumpolar scenes are being used as a basis for algorithm development and testing. Numerous spectral features are being tested and include the 7 basic Landsat TM bands, in addition to ratios, differences, arctans, and normalized differences of each combination of bands. The classifier of choice for the cloud mask is a two-stage approach. In the first stage the class ambiguity for each pixel feature vector is reduced from 10 classes to less than 10 classes through the use of some simple adaptive and fixed thresholds. The feature vector is then passed on to the second stage consisting of a set of specially trained backpropagation neural networks. Each network is trained to resolve specific ambiguities and is selected accordingly. The algorithm is currently being prototyped in C. A Post Launch Operational Plan has also been developed which outlines the process for adapting the current algorithm (designed for Landsat TM) to the ASTER data stream. There are some significant differences between Landsat TM and ASTER and the algorithm will need to be adapted to the additional ASTER channels not present in Landsat TM. A validation plan has also been developed and includes 26 surface validation sites.