CLOUD CLASSIFICATION FOR NOAA-AVHRR DATA BY USING A NEURAL NETWORK

Sample database of clouds, land and water was built based on NOAA-AVH RR 5-channel data which in-clude more than thirty thousand 8×8 pixel samples and more than twenty thousand single pixel samples. The five AVHRR channels are in the at mospheric windows which have a little absorptive mass and are rather trans-parent. They can show characters of detected surface well. Channel 1 (CH 1) is in the visible wave range, CH2 in the near infrared wave range, and CH3, CH4, CH5 are in the inf rared wave range. Theoretical analyses and experiments show t hat not only 5-channel dat a can be used to distinguish clouds, land and water, but also the difference between channels can do so. For example, CH4-CH5 can be used to distinguish water particle cloud and ice particle cloud because the big gest absorption difference between water particle and ice particle is near 12 L m. On the basis of theoretical analyses and experiments, 80 features were extracted from 5-channel AVHRR data for 8×8 pixel samples, which involve spectrum feature, gray feature, channel difference feature, the gray statistical feature and the gray histogram statistical feature. 20 features were selected by using the distinguishing analysis step-by-step method. Classification experiment of sample database was done by using neural net work method. We designed a neural network model with 20 inputs, 2 hidden layers and 4 out puts(20-40-15-4). More then three thousand samples selected randomly were used to train the neural net work model. The other independent samples were used to test. And we got at esting accuracy of 79%. Similarly, for single pixel samples, 20 features including spectrum feature, gray feature and channel difference feature were extracted from 5-channel AVHRR data. Classification ex periment of sample database was made by using neural net work method. We desig ned a neural net work model with 20 inputs, 2 hidden layers and 4 out puts(20-40-15-4). More than two thousand samples selected randomly were used to train the neural network model. The ot her independent samples were used to test. And we gota testing accuracy of 78%. Some different disposals were used to 8×8 pixel samples database and single pixel sample database.