Soft Computing: Frontiers? A Case Study of Hyper-Spectral Satellite Imagin

Soft computing methods such as fuzzy control, neural networks, etc., often require lots of computations even for small amounts of data. It is, therefore, sometimes believed that for larger amounts of data, the required amount of computations will be so large that we will reach the frontiers of soft computing. In this paper, we show, on the example of hyperspectral satellite imaging, that this belief is often too pessimistic. We should not be afraid to use (or at least to try to use) soft computing methods even for large amounts of data. The problem: it looks like soft computing is approaching its frontiers Often, soft computing requires lots of computations. Soft computing methods such as fuzzy control, neural networks, etc., often require lots of computations even for small amounts of data: ̄ When we use fuzzy control to describe a system with n input variables xl ..... xn, then, even if we only use 2 different levels of each variable, we will still need 2n rules. Even for reasonably small n, this is a huge number. ̄ Neural networks are also known to be slow to learn, even for small amounts of data. It is typical to have several thousand iterations to learn a simple dependence. Pessimistic conclusions. If we simply extrapolation this already large amount of computation to the case when we have more input data, we will have to conclude that the required amount of computations will be so large that we will, very soon, reach the frontiers of soft computing. What we are planning to do. In this paper, we show, on the example of hyper-spectral satellite imaging, that this belief is often too pessimistic. We should not be afraid to use (or at least to try to use) soft computing methods even for large amounts of data.