The main approach to identifying coherent structures in a flow field is the Proper Orthogonal Decomposition, due to its simplicity and effectiveness. However it is a data intensive method which becomes more expensive as the data field increases in size. The difficulty pertains mostly when a three-dimensional decomposition is performed, and the limiting factor is storing and loading large data fields of up to billions of gridpoints. This restriction is a conseuqence of the fact that the I/O bandwidth of supercomputers has not been at the same developmental pace as the CPUs. Lossy compression can reduce the size of the data fields and accelerate the computations. The strategy we suggest here relies on data compression via Discrete Chebyshev Transform (or alternatively Discrete Legendre Transform) which leaves invariant the auto-correlation matrix which lies at the core of the POD method. We show that by discarding over 90% of the data we can still retrieve a good proper ortohonal basis of the data set which deviates from the original by 10-2 in the L2 norm.