Richard L. Bankert
Naval Research Laboratory, Monterey, California
Journal of Applied Meteorology
Vol 33, August 1994
pp 909-917
Using advanced very high resolution radiometer data, 16 pixel x 16 pixel sample areas are classified into one of ten output classes using a probabilistic neural network (PNN). The ten classes are cirrus, cirrocumulus, cirrostratus, altostratus, nimbostratus, stratocumulus, stratus, cumulus, cumulonimbus, and clear. Over 200 features drawn form spectral, textural and physical measure are computed from the pixel data for each sample area. The input patterns presented to the neural network are a subset of these features selected by a routine that indicates the discriminatory potential of each feature.
The training and testing input data used by the PNN are obtained from 95 expertly labeled images taken from seven maritime regions; these images provide 1633 sample areas. Theoretical accuracy of the PNN classifier is determined using two methods. In the hold-one-out method, the network is trained on all data samples minus one and is tested on the remaining sample. Using this technique, 79.8% of the samples are classified correctly. a bootstrap method of 100 randomly determined sample sets produces an average overall accuracy of 77.1%, with a standard deviation of 1.4%. In a more general classification using five classes (low clouds, altostratus, high clouds, precipitating clouds, and clear), 91.2% of the samples are accurately classified. A two-layer, four-network system that determines the general classification of a sample followed by a specific classification in another network is proposed. Testing of this system produces mixed results compared to the single ten-class PNN.
With the recent success of a neural network approach for the classification of land, water, and sky elements in polar images (Welch et al. 1992), a similar investigation is performed using a probabilistic neural network (PNN) to classify components of images from Northern Hemisphere maritime regions. Spectral, textural, and physical features are computed from AVHRR LAC channels 1 and 4 data to classify ten cloud types (including clear).
After a feature selection procedure, two test are applied to the data to estimate the theoretical accuracy of the PNN classifier. In the hold-one-out method, 79.8% of the samples are correctly classified. The overall average accuracy for the modified bootstrap method is 77.1%, with a standard deviation of 1.4%. The pattern of misclassifications for the various cloud types is similar for both methods.
Other test performed include using only special features. The results have reinforced the importance of using a combination of feature types. When using only Atlantic Ocean data in another test, there is only a slight change in both overall and class accuracies compared to the results when the remaining data from the other six regions are added. Thus, there appears to be no geographical bias to these results. In another test, Cirrus and Cirrostratus samples are combined into one class. Test accuracies are approximately 4% higher than the original ten-class PNN.
A general classification is performed that reduces the number of classes to five. The PNN correctly classified 91.2% of the samples during a hold-one-out test. A major error occurs with the misclassification of Altostratus samples as low clouds.
A two-layer, four-network system that allows a sample to be first classified in the five-class PNN followed by a more specific cloud-type classification is proposed. The system has the advantage of using features that are unique to the separation of a smaller set of classes. These features may not have been selected as significant discriminates in the ten-class problem. Their inclusion was expected to increase overall classification accuracy with a better discrimination of difficult classes. However, the class accuracy results achieved in the approach are mixed in comparison with the original ten-class PNN results.
Data problem areas that may have had an influence on all of the accuracy test include background contamination and mixing of image samples with different resolution. Not all of the 16 X 16 sample regions are completely filled with the cloud type for which they are labeled; therefore, there could be some distortion to the computed features with the mixture of cloud and background pixels. DIfferent resolutions occur since samples that are farther away from nadir have a lower resolution, which would alter the pixel-by-pixel representation of any particular cloud type. The consistency of the features for that class may be lessened.
Introducing features from AVHRR channel 3 and other data sources may improve the classification accuracy. AVHRR channel 3 (near IR) data tend to be noisy such that images would have to be put through a noise reduction filter before features could be extracted. The enhanced images resulting from data operations (subtraction, division, etc.) among channels may provide helpful features. There is also the possibility of adding upper-air data from computer models or other sources. This information might lead to a relation between the cloud class and the temperature or dewpoint sounding or other information that could be extracted from upper-air data.
In comparison with the spectral-only classifier, the PNN's ability to separate the ten classes is enhanced with the addition of textural and physical features. However, there is some indication that other discriminating features (like those discussed above) should at least be examined to improve the separation of the more problematic cases (Cirrus vs. Cirrostratus, Stratus vs. Stratocumulus, and Cumulus vs. Stratocumulus). Being able to discriminate these classes will improve the accuracy substantially.
All of the test results presented here are encouraging, but are the theoretical accuracies acceptable for operational use? Any potential user of a cloud classifier has to make a determination depending upon the need for detail and the level of confidence required. If an application of a resultant cloud classification requires a high level of detail (e.g., separation of Cirrus and Cirrostratus) with a high level of confidence, the current ten-class PNN classifier probably would not be satisfactory. However, for less demanding applications, where the more probable misclassifications (Cirrus vs. Cirrostratus, Stratus vs. Stratocumulus) are not considered a serious drawback, the ten-class PNN would probably provide an acceptable classification. Alternatively, a user may want to consider using the five-class PNN in situations that require a highly accurate (>90%), albeit more general, classification.