A NEURAL NETWORK APPROACH TO CLOUD CLASSIFICATION

Recent cloud retrieval validation studies suggest that there are major discrepancies between the various algorithms. Other studies have demonstrated that the use of texture-based pattern recognition features can signifcantly improve cloud identifcation accuracy. However, the capabilities and accuracies which can be attained with spatial information remain poorly understood and undocumented, and the choice of an optimal feature set is unknown. The results from this study demonstrate that, using high spatial resolution data, very high cloud classifcation accuracies can be obtained. A texture-based neural network classifer using only single-channel visible LANDSAT MSS imagery achieves an overall cloud identifcation accuracy of 93%. 11 is remarkable that cirrus can be distinguished from boundary layer cloudiness with an accuracy of 96%, without the use of an infrared channel. Stratocumulus is retrieved with an accuracy of 92%, cumulus at 90%. The use of the neural network does not improve cirrus classifcation accuracy. Rather, its main effect is in the improved separation between stratocumulus and cumulus cloudiness. While most cloud classifcation algorithms rely on linear, parametric schemes, the present study is based on a nonlinear, nonparametric four-layer neural network approach. Intercomparisons are made to a three-layer neural network architecture, the nonparametric K-nearest neighbor approach, and the linear stepwise discriminant analysis procedure. A signifcant finding is that signifcantly higher accuracies are attained with the nonparametric approaches using only 20% of the database as training data, compared to 67% of the database in the linear approach. © 1990 IEEE