| Tropical Weather States from the ISCCP D1 Cloud Dataset | ||
|
Introduction The International Satellite Cloud Climatology Project (ISCCP) D1 gridded cloud product contains the distribution of cloud top pressure and cloud optical depth at each grid cell observed every 3-hours daily from 1984-2004. The distribution of cloud top pressure (PC) and optical depth (&tau) fractions is referred to as a PC-tau histogram. An histogram displaying cloud top pressure and optical depth over some region reveals information regarding the type of clouds present (Rossow and Schiffer, 1999). Each PC-tau histogram contains 42 boxes, 6 optical thickness bins and 7 cloud top pressure bins (the bin sizes are not linear). One D1 PC-&tau histogram contains the relative percentage of pixels that fall into each of these 42 boxes in the corresponding grid box (2.5°x2.5°). We extend to the whole tropics (±15° latitude) the analysis approach of Jakob and Tselioudis (2003, henceforth JT03) that uses the patterns of cloud property joint distributions at mesoscale (cloud type mixtures) to identify distinct states of the tropical atmosphere. This approach provides a more compact way to study multi-variate relationships among clouds, meteorology and energy-water exchanges (Rossow et al, 2005).
Data and Analysis Method
We look for distinctive patterns in the joint frequency distributions of the cloud top pressure (PC) and optical thickness (&tau) values from individual satellite image pixels (fields-of-view about 5 km in size) occurring within 2.5° regions (upper limit of the mesoscale) that are provided in the ISCCP D1 dataset (Rossow and Schiffer 1999). The histogram patterns that describe cloud variability are identified using the K-Means clustering algorithm (Anderberg, 1973) applied to 3-hourly PC-&tau histograms for each 2.5° region, including completely clear regions. To download the analysis software, click HERE. Unlike JT03, the "best" number of clusters is determined objectively by repeating the analysis for an increasing number of clusters (starting at four, since JT03 tested 2-4 clusters) and judging the outcome by four criteria: (1) the resulting centroid histogram patterns must not change significantly (as judged by the pattern correlations among the centroids) when the random number initiating the analysis or the subset of data analyzed is changed, (2) the resulting centroid patterns should differ from each other significantly (pattern correlations should be low, usually < 0.6), (3) the spatial-temporal correlations of the centroid histograms should also be low, and (4) the distance between cluster centroids should be larger than the dispersions of the cluster member distances from the centroid. Tests for K=5-7 showed that the results were unstable (violation of the first criterion) when the cluster number was too small (K < 6 in this case) and that, when the cluster number was increased, the patterns of the new clusters were still significantly different from each other (second criterion). When there were too many clusters (K > 6), two or more of them would have very highly correlated PC-TAU patterns, as well as highly correlated spatial distributions or time variability (usually all three). In particular, two of the K-cluster patterns are highly correlated with a single cluster pattern from the (K-1) analysis, indicating a splitting of that cluster. The last criterion is actually used in the analysis to optimize the cluster set, so it is always met; but a post-facto check showed generally decreasing dispersion-to-separation distance ratios as the number of clusters increases. For K = 6 the average dispersion-to-separation distance ratio is less than one half. The optimum cluster centroids are identified for K=6 as shown above and reported below on the left. Next, each PC-TAU histogram for each 3-hr time step in each 2.5° map grid cell over the whole tropics is assigned to one of these clusters. We refer to these as "weather states" (WS) because JT03 showed that these cloud property patterns are associated with distinct states of the tropical atmosphere (Fu et al. 1990, 1994, Lau and Crane 1995, Del Genio and Kovari 2002). Variations of WS can then be represented by the frequency of occurrence maps for daily or longer time accumulations shown below.
Note: Weather state values range from 1-6. In the cluster datasets, the value 0 indicates Clear; value 98 indicates darkness; and value 99 indicates missing data. To download the dataset (D1.WS.Tropics.1.dat), To download the cluster centroid (D1_TR_all.cent6.3.dat)click HERE . To download the routine converting those datasets in binary files, click HERE. PLEASE NOTE -- An error in the first version of the Tropical cluster dataset, which caused data from 3/1/1988-12/31/1988 to be repeated from 3/1/1989-12/31/1989, was discovered. The dataset has been replaced. The current version is D1.WS.Tropics.1.dat. More information are available on the ISCCP website ... just follow this LINK !
References
|
