The problem of producing polar cloud masks for satellite imagery is an important facet of the research on global warming. For the past three years, our research on this topic has produced a series of classifiers. The first classifier used traditional statistical techniques, and, although the performance was reasonably good, better accuracy and faster classification speeds were desired. Neural network classifiers provided an improvement in both classification speed and accuracy but a single monolithic network proved difficult to train and was computationally expensive. A decomposition of the neural network into a hierarchical structure provided significant reductions in training time and some increase in accuracy. While this technique produced excellent results, to optimize its performance a minimal feature set and a highly accurate and easily computed switching mechanism must be identified. This paper presents recent developments in these two areas. Landsat Thematic Mapper (TM) data from the arctic and antarctic was used to test the network. A minimal feature set, which defines the elements of the network input vector, is desirable for both improving accuracy and reducing computation. A smaller input vector will reduce the number of weights which must be updated during training and concomitantly reduce training and testing times. Small input vectors are also desirable because of the oft-cited 'curse of dimensionality' which states the higher the dimension of the problem to be solved, the more difficult it will be for the network to find an acceptable solution. However, it is also known that if a network has insufficient information, it will not be possible to form an appropriate decision surface. In that case, additional features, and additional dimensions, are required. Finding the proper balance can be difficult. Previously, trial and error was used to find a 'good' selection of features for classification. Features were added individually and those which had no impact on the neural network classification were deleted. A new approach clusters the features and selects the members of the cluster with the highest information content. Features are still removed one at a time but clustering ensures that statistically different features are preserved and allows for parallelization of the feature selection process. In addition, a fuzzy expert system provides a much faster classification accuracy approximation. This technique was able to significantly reduce the size of the feature vectors without sacrificing classification accuracy. The switching mechanism used in this work employs a series of static and adaptive thresholds derived from statistical analysis of polar scenes. This technique is faster than the corresponding neural network switching mechanism and can be easily changed as additional data becomes available. The resulting system, then, uses the adaptive thresholds to select the appropriate neural network from a collection of multilayer perceptron networks each responsible for classifying a subset of the total number of classes. The inputs to these networks are selected by the fuzzy logic algorithm. The difficulty of finding these thresholds, a task performed by a human expert, motivated the use of a genetic algorithm to determine these values. This system was able to achieve 96.45% accuracy on the fundamental problem of distinguishing cloud from non-cloud classes. The time required to classify 468,750 pixels in a satellite image was 50 seconds.3220