No abstract

No abstract

No abstract

The previous cloud classification investigations in parts 1–3 of this study have been conducted using 1/16‐km Landsat Multispectral Scanner (MSS) imagery. However, for global monitoring, much lower spatial resolutions of the order of 1–8 km generally are used. The present study examines the loss of cloud classification accuracy as a function of spatial resolution by degrading the imagery through progressive averaging. Textural measures are computed using the Gray Level Difference Vector approach. Significant improvement in cloud classification accuracy can be obtained using 1/2‐km spatial resolution data rather than the current 1‐km resolution data available today from AVHRR and GOES. Cirrus classification accuracy is especially compromised as the spatial resolution is degraded. However, the use of texture measures defined at the combination of pixel separations d = 1, 4 improves classification accuracies by several percent even for 1‐km spatial resolution data. Cirrus accuracy is significantly improved by use of multiple distance features. Classification accuracies using 1/8‐km spatial resolution data are similar to those obtained using the full spatial resolution features. The implications are that there are no advantages to be gained in cloud classification accuracies by using even higher spatial resolutions available from Landsat Thematic Mapper or SPOT imagery. Finally, it is found that multiple resolution imagery can be used to improve classification accuracy. Indeed, the “global” rather than the “local” aspects of texture appear to be most important to the classifier.

It is widely recognized that cloud classification schemes based upon multispectral signatures and clustering measures are severely limited over snow-and ice-covered surfaces. This is due to the similarity of cloud and snow/ice spectral signatures in both visible and infrared wavelengths. Infrared threshold techniques are limited, in particular, by persistent surface inversions and warm low-level clouds. However, pattern recognition schemes based upon the combination of spectral and textural signatures can be used effectively for cloud discrimination over high albedo surfaces.This study is based primarily upon AVHRR LAC imagery, but with some results from LANDSAT high spatial resolution scenes. A large number of textural features are investigated, including the Gray Level Difference Vector (GLDV) and Sum and Difference Histogram (SADH) approaches, various features proposed by Garand,1 the Gray Level Run Length (GLRL), spatial coherence "footprints,' and spectral histogram measures. Twenty arctic surface and cloud classes are identified using two different classificatiofl approaches: 1) the traditional stepwise discrirninant analysis and 2) neural network analysis. Principal component analysis of textural measures is used to eliminate those measures which contribute little to class separability. The neural network feed-forward, back-propagation approach produces the highest classification accuracy, and does so with a relatively small training set. However, the main limitation is the long training times required. A new hybrid architecture using a modularized competitive learning layer inserted before the feed-forward backpropagation layer developed by Lee and Weger2 also is being included. Preliminary results suggest comparable cloud classification accuracy with a training time two orders of magnitude faster.