Skip to Main Content
Automatic retrieval and interpretation of satellite images is critical for managing the enormous volume of environmental remote sensing data available today. It is particularly useful in oceanography and climate studies for examination of the spatio-temporal evolution of mesoscalar ocean structures appearing in the satellite images taken by visible, infrared, and radar sensors. This is because they change so quickly and several images of the same place can be acquired at different times within the same day. This paper describes the use of filter measures and the Bayesian networks to reduce the number of irrelevant features necessary for ocean structure recognition in satellite images, thereby improving the overall interpretation system performance and reducing the computational time. We present our results for the National Oceanographic and Atmospheric Administration satellite Advanced Very High Resolution Radiometer (AVHRR) images. We have automatically detected and located mesoscale ocean phenomena of interest in our study area (North-East Atlantic and the Mediterranean), such as upwellings, eddies, and island wakes, using an automatic selection methodology which reduces the features used for description by about 80%. Finally, Bayesian network classifiers are used to assess classification quality. Knowledge about these structures is represented with numeric and nonnumeric features.