COMPARISON OF BACKPROPAGATION, CASCADE-CORRELATION AND KOHONEN ALGORITHMS FOR CLOUD RETRIEVAL

The expected high volume imagery data from NASA Mission to the Planet Earth (e.g. one terabyte for EOS AM) is one of the target application areas for automated cloud retrieval, and more generally for automated image classification. We used the backpropagation (BP), the Cascade-Correlation (CC) and Kohonen self-organizing map (SOM) neural network architectures for cloud retrieval from satellite imagery. We have used a simple scene (a mixed scene containing only cloud and ocean). This simple scene allows us to evaluate the accuracy of the classification (and the trend in misclassifications). better than a complicated scene. BP and CC performed at the same accuracy level. CC was slightly more efficient than BP in terms of the number epochs. BP requires the user to set the number of hidden units. CC demonstrated a built -in flexibility in terms of the (variable) number of hidden units necessary to accomplish the learning phase of the algorithm. The SOM algorithm, was slightly less accurate, do to its unsupervised nature, than CC and BP for our test data ( 97% accuracy versus 99% accuracy level for BP and CC). This study shows that for simple scenes, which are abundant in global monitoring satellite imagery, a simple pixel-by-pixel or 3-by-3 window approaches provide high accuracy classification without using complicated contextual information.