Hybrid neural network system for cloud classification from satellite images

This paper presents an architecture and preliminary implementation results of a hybrid two-stage neural network system for cloud classification from satellite imagery. The system first performs pixel classification on the image spectral multi-channel data and descriptive data to discover possible areas covered by clouds and cloud contaminated pixel characteristics. Then it investigates the texture of image rectangular kernels composed of classified pixels belonging to classes recorded previously with some expected to represent clouds. The system determines cloud textures, integrates pixel information from within local image areas, and provides the final cloud classification. The method is based on the unsupervised classification approach. The hybrid neural network used consists of classical and modified learning multilayer self-organizing feature maps. The preliminary tests have been made on both artificial and satellite image data. The initial results are satisfactory and promising.