A neural network approach to cloud detection in AVHRR images

Abstract

The problem of identifying clouds and fog areas from advanced very high resolution radiometer (AVHRR) images using a neural network approach is used. The backpropagation paradigm was used to train many different architectural configurations of the neural network to classify the cloud content of an 8*8-pixel window in an image into five categories (ranging from 100% cloudy to 0% cloudy). The results indicate a large range in the performance of the different architectures. The most successful architectural configuration was used to create cloud masks for a series of AVHRR images. The cloud masks generated compared favorably with a trained analyst and other automated techniques.< >