Improvement and Application of Artificial Neural Networks to Cloud Classification

2449 cloud classification samples are artificially selected from infrared channel 1, infrared channel 2 and water vapor channel of VISSR on FY-2C geostationary meteorological satellite during 2005—2009. Different linear combination of three channels are selected as feature values, which are brightness temperature of IR1, IR2, WV, and brightness temperature difference of IR1 to IR2, IR1 to WV, IR2 to WV. According to statistical theory, the sample probability distribution is assumed to help to remove some apparent unreasonable data such as outliers, and to understand cloud normal features better. It is found that the brightness temperature difference of IR1 to IR2 is most sensitive to the amount of thin cirrus cloud in 6 selected values. On the other hand, the error of the BP neural network model mostly comes from the contradiction of this feature too. A nested BP artificial neural network model is designed, and it's composed of two layers. The first layer includes five features of brightness temperature of IR1, IR2, WV, and brightness temperature difference of IR1 to WV, IR2 to WV that are used to classify each pixel to one of four categories such as clear, mixed cloud, thick cirrus cloud and strong convective cloud. And the second layer includes just one feature, brightness temperature difference of IR1 to IR2 that are used to classify mixed cloud to low-level cloud, mid-level cloud or thin cirrus cloud. Finally, every pixel is classified into one of total 6 categories corresponding to each color. Both layers adopt a BP neural network, the most widely used algorithm for generating classifiers, with one hidden layer and the additional momentum method, not only accelerating the training speed, but also reducing the redundancy of the networks.Error analysis shows that the accuracy rates of the nested BP artificial neural network for types of mid-level cloud and thin cirrus cloud have increased by 42.6% and 11.3%, respectively. The mean square error and normalized mean square error of the whole classification model have decreased by 6.1% and 44.7% with the correlation coefficient increasing by 3.4%. By comparison of the classification results from 3 tests of tropical, subtropical areas and tropical cyclone, it shows that the nested model identifies thin cirrus cloud more accurately than the traditional model. Therefore, the results of the nested model are more reasonable than the traditional model.