Cloud field classification based upon high spatial resolution textural features: 1. Gray level co‐occurrence matrix approach

Abstract

Standard cloud algorithms rely on multispectral signatures to identify high, medium, and low clouds. In contrast, the present study classifies stratocumulus, cumulus, and cirrus clouds, using textural features alone, derived from a single high‐resolution Landsat Multispectral Scanner near‐infrared channel. Applying stepwise linear discriminant analysis, classification accuracies of 92–95%, with standard deviations of about 1.5%, are achieved using the random subregion hold‐out pattern. Accuracies of 83–88%, with standard deviations of about 7%, are achieved using the random scene hold‐out pattern. The theoretical accuracy of the approach is estimated by treating the simulation on a non‐parametric basis, using the bootstrap method. It is significant that the present method is capable of distinguishing high cirrus clouds from low clouds strictly on the basis of spatial brightness patterns. In fact, cirrus has the highest classification accuracy with this approach. The largest probability of misclassification is associated with confusion between stratocumulus breakup regions and fair‐weather cumulus. The present study is based upon textural features computed from Gray Level Co‐occurrence Matrix (GLCM) statistics defined at various pixel displacement distances. It is found that textural features defined at pixel separations of 0.5 km produce classification accuracies equal to or better than those defined at separations of 57 m. Some slight improvement (2–3%) in classification accuracy is achieved by inclusion of textural features defined at multiple pixel separations. In general, stratocumulus classification is improved by inclusion of additional features defined at pixel separations d < 0.25 km, cirrus by features at d ≃ 1 km, and cumulus by features at d ≳ 2 km.

Temporal updating scheme for probabilistic neural network with application to satellite cloud classification

“…In the summation layer, the weight from th neuron in pool of pattern layer to the th neuron in the summation layer is . By configuring the PNN in this way, the output of the PNN will be the same as the Gaussian mixture model output given by (4) and (5). Since generally is much smaller than the number of training samples that belong to class , , the pattern layer of the PNN is therefore substantially simplified from its original version.…”

“…The Gaussian mixture model described in (4) and (5) can also be easily mapped to the PNN structure. For the th pool in the pattern layer, only neurons are needed.…”

“…For any class , suppose the class conditional distribution is modeled approximately by a Gaussian mixture, i.e., (4) where is the number of Gaussian components in class and 's are the weights of the components which satisfy the constraint , denotes the multivariate Gaussian density function of the th component in class . Further, we have…”

“…Textural features, which are often distinct and tend to be less sensitive to the effects of atmospheric attenuation or detector noise, have received more attention in recent years. Welch et al [4] calculated several statistical textural features based on gray level co-occurrence matrix (GLCM) while Lamei et al [5] used Gabor filters to extract cloud textural features. Several comparative studies of both the spectral and textural features for cloud and other satellite imagery classifications have been conducted by Parikh [6], Gu [7] and Ohanian [8] and Augusteijin [9] which led to the conclusion that no consistent optimal feature extraction scheme can be devised for this problem.…”

“…Simpson and Gobat [10], [11] used a nested hierarchical partitional clustering algorithm to segment GOES images and an adaptive thresholding to label clusters as cloudy or cloud free classes. Welch [4] used linear discrimination techniques while Lee et al [12] tested a three-layer back-propagation neural network (BPNN). A Probabilistic neural network (PNN) was also examined by Bankert et al [13].…”