One method I have used for sparse data,ie data with a lot of null readings is to pass the data serialy through a zero counter which increments when it sees a zero and dumps the resultant count out when it sees any other value,any other values are written out as thay are.It workswell on sparse data as its a kind of sort of modified huffman coding system.If your data isnt sparce then you needto do some more eloborate huffman coding system with symbols for a hierachy of increasing entropy.