A System Based on Ratio Images and Quick Probabilistic Neural Network for Continuous Cloud Classification

An approach for continuous daytime cloud classification system through satellite images is presented. The system is based on spectral ratio values as input features and a modified version of probabilistic neural network (PNN), named Quick PNN (QPNN), as a classifier. The use of spectral ratio values makes the system more efficient in detecting the minor changes in cloud spectral properties, leading to better classification capability. The modification to PNN consists of shrinking the hidden layer which is accomplished by performing K-means clustering on the training data of each class separately. Thus, for each class, instead of presenting all the training data samples in the hidden layer nodes, only the means of the resultant clusters are presented. The training data and the class labels are derived through the generation and interpretation of ratio images. The application of the approach to Meteosat-8 images has shown the separation of eight classes, including low clouds, middle clouds, high clouds, areas of high water vapor, sea surface, and land. The average accuracy of the system is 87.15% with a range of 84%-91% for the cloud and area of high water vapor classes, 93% for sea surface class, and 85% for land surface class. The computation time of the classification mode, including image ratioing and QPNN operations, is less than 1 min, which is good for continuous cloud classification and monitoring. The approach can be adapted to any multichannel satellite sensor only by using proper combination of ratio images.