The BDS contains the packed data and the binary scaling information needed to reconstruct the original data from the packed data. The required decimal scale factor is found in the PDS, above. The data stream is zero filled to an even number of octets. 
Octet no.  BMS Content 

Length in octets of binary data section  
Bits 1 through 4: Flag
(See Table 11) Bits 5 through 8: Number of unused bits at end of Section 4.  
The binary scale factor (E). A negative value is indicated by setting the high order bit (bit No. 1) in octet 5 to 1 (on).  
Reference value (minimum value); floating point representation of the number.  
Number of bits into which a datum point is packed  
Variable, depending on octet 4; zero filled to an even number of octets.  
Optionally, may contain an extension of the flags in octet 4. (see Table 11) 
Here are some of the various forms the binary data can take; the flag table in BDS octet 4, possibly extended into octet 14, identifies which variant is in use. Gridpoint data  Simple packing Here the data simply begin in octet 12 and continue, packed according to the simple packing algorithm described above, without any particular regard for computer "word" boundaries, until there is no more data. There may be some "zerofill" bits at the end. If all the data in a grid point field happen to have the same value, then all of the deviations from the reference value are set to zero. Since a zero value requires no bits for packing, octet 11 is set to zero, thus indicating a field of constant data, the value of which is given by the reference value. Under these circumstances, octet 12 is set to zero (the required "zero fill to an even number of octets") and bits 58 of octet 4 contain an 8. The number of data points in the field is implied by the grid identification given in the PDS and/or the GDS and BMS. Spherical Harmonic Coefficients  Simple packing Octets 1215 contain the real part of the (0.0) coefficient in the same floating point format as the reference value in octets 710. The imaginary part of the (0.0) coefficient, mathematically, is always equal zero. Octets 16 to the end contain the remaining coefficients packed up as binary data with the same sort of scaling, reference value, and the like, as with gridpoint numbers. Excluding the (0,0) coefficient, which is usually much larger than the others, from the packing operation means that the remaining coefficients can be packed to a given precision more efficiently (fewer bits per word) than would be the case otherwise. GridPoint Data  Second Order or Complex Packing Before laying out where the various second order values, subparameters, counters, and what have you, go, it is appropriate to describe the second order packing method in an algorithmic manner. Referring back to the description of simple packing, the encoding method is the same up to part way through the fourth step, stopping just short of the actual packing of the scaled integers into the "words" of either a prespecified or calculated bit length. The basic outline of second order packing is to scan through the array of integers (one per grid point, or possibly less than that if the Bit Map Section has been employed to discard some of the null value points) and seek out subsections exhibiting relatively low variability within the subsection. One then finds the (local) minimum value in that subsection and subtracts it from the ("first order") integers in that subsection, which leave a set of "second order" integers. These numbers are then scanned to find the maximum value, which in turn is used to specify the minimum bit width for a "word" necessary to contain the subsection set of second order numbers. The term "first order" in this context refers to the integer variables that result from subtracting the overall (global) minimum from the original variables and then doing all scaling and rounding; "second order" refers to the variables that result from subtracting the local minimum from the subset of first order variables. No further scaling is necessary or appropriate. The subsection set of numbers are then packed into "words" of the just determined bit length. The overall savings in space comes about because the second order values are, usually, smaller than their first order counterparts. They have, after all, had two minima subtracted from the original values, the overall minimum and the local minimum, where the first order values have had only the overall minimum subtracted out. There is no guarantee, however, that the second order packing will compress a given field to a greater degree than the first order packing. If the first order field of integers is highly variable, or generally close to zero, then there will be no gain in compression. But if the field shows long runs of small variation, particularly if some of the runs are constant (zero variability), then the second order packing will contribute to the compression. The process then repeats and a whole collection of subsections are found, their local minima are subtracted, etc. One of the tricky parts of this process is defining just what is meant by a "subsection of low variability". The WMO Manual is silent on this as it only describes how the subsections and their ancillary data are to be packed in the message. The U.S. National Weather Service, the U.K. Meteorological Office, the European Centre for MediumRange Weather Forecasts, and probably other groups have, independently, designed selection criteria and built them into GRIB encoders. It is beyond the scope of this document to attempt to describe them in any detail. These groups have all expressed their willingness to share their GRIB encoders with any who ask for them. Before laying out where the second order values, etc., are placed in a message, we had best review just what information has to be saved. We need to include the following information:
A moments consideration (a long moment, perhaps) will satisfy the reader that the information given will be sufficient to reconstruct the original data field. The information needed for points 2) and 3), the beginning and end of the subsections, is presented in the form of a bit map, called a "secondary bit map" to distinguish it from the bit map (optionally) contained in the BMS. There is one bit for each grid point containing data, ordered in the same way as the grid is laid out. The "primary" bit map, the BMS bit map, may have been used to eliminate data at points where the data are meaningless  only the remaining "real" data points are matched by the bits in the secondary bit map. This possibility is understood to exist throughout the following discussion. The start of each subsection is indicated by the corresponding bit set to "on" or to a value of 1. Clearly, the first bit in the secondary bit map will always be set on, since the first data point must be the start of the first subsection. (If it is not, then something is wrong somewhere. Unfortunately it is not always easy to tell just where the error occurred.) The secondary bit map is then no more than a collection of 1s and 0s, indicating the start and the extent of each subsection. It would be possible to scan through the secondary bit map and determine how many subsections there are; however, this number is explicitly included in the GRIB message to save one the trouble, and to serve as an internal selfchecking mechanism. At long last, then, here is the layout of the information, with further explanatory notes, when second order packing has been employed:
There are a small number of special cases and variations on the above layout:
