Home
Mining Environment
Research Projects
Education
Case Studies
Tools
publications
Staff
ITSC
Contact Us

 

NASA logo

ITSC logo
Information Technology and Systems Center The University of Alabama in Huntsville
Huntsville, AL 35899 (256) 824-6868
info@itsc.uah.edu

UAH logo

 

Detecting Cumulus Cloud Fields in GOES Images
Texture features commonly have been used in remote sensing applications such as land classification and detection of atmospheric phenomena. Recently, atmospheric scientists have become interested in the effects of land use changes such as urbanization and deforestation on climate. The study of the formation of cloud fields is an important problem in this research area. The GOES satellites are geosynchronous satellites used for monitoring and predicting weather. Cloud fields often begin as boundary layer cumulus fields consisting of many small scattered clouds. Cloud detection algorithms based on spectral properties of the data often miss these clouds. This is due to the fact that such clouds can have a diameter of 1km or less, which is significantly less than the size of a single infrared pixel. In order to detect boundary layer cumulus clouds, the higher resolution visible channel must be used. Identification of cumulus cloud fields cannot be based on a single pixel value, since the intensity of the pixels in the field varies dramatically. Area properties such as texture are therefore ideal for identifying cumulus cloud fields.

Three different texture based techniques (Gray Level Run Length (GLRL), Gray Level Cooccurence Matrix (GLCM) and Association Rules) were compared for detecting cumulus cloud fields. A training phase is required for the selection of texture features and for training the classifiers. A total of one hundred sample images of size 13 X 13 are extracted. Fifty images are used for training the classifier and the remaining images are used for testing purposes. Of the fifty training samples, ten represent the background class, twenty represent the cumulus cloud class, and twenty represent the other cloud class. The test samples are chosen in the same way. The training and test samples are drawn from different images. The performance of the three features on the test samples is shown in the table below.


Classification performance on GOES test samples

  Association Rule GLCM GLRL
Background 10/10 = 100.0% 10/10=100.0% 10/10 = 100.0%
Cumulus 19/20 = 95.00% 16/20=80.00% 19/20 = 95.00%
Other Clouds 20/20 = 100.0% 20/20=100.0% 100.0%
Overall Accuracy 98.00% 92.00% 98.00%

Segmentation accuracy is measured using images labeled or segmented by atmospheric science experts. Six scenes of size 512 X 512 pixels are labeled by two independent experts. Due to the difficulties described previously, some pixels are labeled differently by different experts. Only those pixels that receive consistent labeling are considered in the accuracy computation. These scenes are segmented using association rule, GLCM and GLRL based features. Association rule features yield an average accuracy of 89.11% on the six scenes, compared to 72.98% for GLCM features and 88.26% for GLRL features.

Collaboration with Domain Expert:
U.S. Nair and Dr. Ron Welch (UAH Atmospheric Science)