Permalink
Fetching contributors…
Cannot retrieve contributors at this time
239 lines (200 sloc) 22.1 KB
LASzip for LAS 1.4 (native extension)
========
Specification Document
--------
As before the LASheader and the VLRs are stored uncompressed but the highest bit of the point type field is set (i.e. the compressed point types 128 to 138 correspond to the uncompressed point types 0 to 10). LiDAR points are compressed in completely independent chunks that default to 50,000 points per chunk. Each chunk can be decompressed "on its own" once the start byte is know. For point types 0 to 5 nothing changes and the same compression scheme as before is used. For the new point types 6 to 10 of LAS 1.4 there are two main changes: the chunks are encoded in layers and the chunks can have variable size.
As before the first point of each chunk is stored raw. The attributes of all following points are broken into several layers. Only the first layer containing the x and y coordinates (and a few pieces of information) is mandatory to read and decompress. The remaining layers containing independently useful attributes such as elevation, intensity, classifications, flags, GPS times, colors, point source ID, user data, scan angles, wavepackets, and extra bytes do not necessarily need be decompressed (or be read from disk).
Another new (or rather "revived") feature in LASzip compression for the new point types of LAS 1.4 is that it will be possible to continously vary the chunk size. This was in the design of the original LASzip coder but broke whenever the file was streamed due to missing access to the per chunk point counts during a streaming read as those were only stored in the chunk table at the very end of the LAZ file. This is fixed now (but only for the new point types) by also storing the number of points at the beginning of each chunk.
LAZ File Layout:
----------------
1. LASheader
2. VLRs + LASzip VLR
3. Sequence of n Chunks
1. Chunk 1
2. Chunk 2
3. Chunk 3
....
n. Chunk n
4. Chunk Table (chunk starts and point numbers)
5. EVLRs
6. LASindex EVLR (future feature: optional index for faster spatial queries)
7. LASlayers EVLR (future feature: for in-place edits of classifications / flags)
Chunk Layout:
-------------
1) Raw point [30 - 68 (or more) bytes]
2) Numbers and Bytes
+ Number of *remaining* points [4 bytes]
+ Number of bytes (maybe compressed)
- scanner channel, point source ID change, GPS time change, scan angle change, return counts, and XY layer [4 bytes]
- Z layer [4 bytes]
- classification layer [4 byte]
- flags layer [4 bytes]
- intensity layer [4 bytes]
- scan angle layer [4 bytes]
- user data layer [4 bytes]
- point source ID layer [4bytes]
- GPS time layer [4 bytes]
optional
- RGB layer [4 bytes]
- NIR layer [4 bytes]
- WavePacket layer [4 bytes]
- "Extra Bytes" layer [4 bytes per byte]
3) Layers
- scanner channel, point source ID change, GPS time change, scan angle change, return counts, and XY layer
- Z layer
- classification layer
- flags layer
- intensity layer
- scan angle layer
- user data layer
- point source ID layer
- GPS time layer
optional
- RGB layer
- NIR layer
- WavePacket layer
- "Extra Bytes" layer(s)
Compression of scanner channel, point source ID change, GPS time change, scan angle change, return counts, and XY layer
-----------------------------------------------------------------------------
Due to the new scanner channel it is *crucial* to first encode whether a point is from the same and if not from which other scanner channel so that the correct context can be used for all subsequent predictions. We also encode whether a point has a a different point source ID, a different GPS time, and a different scan angle as the previous point from the same scanner channel as these changes correlate strongly with another. The return counts that are also recorded in this layer.
* scanner channel compared to the scanner channel of the previous point (same = 0 / different = 1)
* point source ID compared to the point source ID of the previous point from the *same* scanner channel (same = 0 / different = 1)
* GPS time stamp compared to the GPS time stamp of the previous point from the *same* scanner channel (same = 0 / different = 1)
* scan angle compared to the scan angle of the previous point from the *same* scanner channel (same = 0 / different = 1)
* number of returns compared to the number of returns of the previous point from the *same* scanner channel (same = 0 / different = 1)
* return number compared to the return number of the previous point from the *same* scanner channel (same = 0 / plus one = 1 / minus one = 2 / other difference = 3)
These 7 bits of information are combined into one symbol whose value ranges from 0 to 127 that we then compress with one of four (4) different contexts based on whether the *directly* previous point (no matter from which scanner channel) was a single return (0), or the first (1), the last (2) or the intermediate (3) return in case of multi-return.
Only if the **scanner channel is different** (as indicated by these 7 bits) we use one symbol whose value ranges from 0 to 2 to encode whether we need to add 1, 2, or 3 to the previous scanner channel to get (modulo 4) to the current scanner channel that we then compress using the previous scanner channel as one of four different contexts. All following predictions in all different layers are relative to the previous point from the *same* scanner channel. They first check whether the current point is from the same scanner channel and switch context if not. If no previous point exists for the new scanner channel (because this scanner channel appears the first time in this chunk) then the previous point (from whichever other scanner channel) is used. For the very first point per chunk *that gets compressed* this is that one point that is stored raw at the beginning of every chunk.
Only if the **number of returns has an other difference** (as indicated by the 7 bits earlier) we use one symbol whose value ranges from 0 to 15 that we then compress with the previous number of returns (from the same scanner channel) as one of sixteen contexts.
Only if the **return number is different** (as indicated by the 7 bits earlier) we encode in in two possible ways depending on whether the GPS time stamp has changed:
- if the GPS time stamp *has not* changed we use one symbol whose value ranges from 0 to 12 to encode whether we need to add 2, 3, 4 ... 12, 13 or 14 to the previous return number (from the same scanner channel) to get (modulo 16) to the current return number that we then compress using the previous return number (from the same scanner channel) as one of sixteen contexts.
- if the GPS time stamp *has* changed we use one symbol whose value ranges from 0 to 15 to encode the current return number that we then compress with the (already encoded) number of returns as one of sixteen contexts.
Next we use the number of returns n and the return number r to to derive two numbers between 0 and 15 that are used for switching contexts later: a *return map m* and *a return level l*.
The *return map m* simply serializes the valid combinations of r and n: for r = 1 and n = 1 it is 0, for r = 1 and n = 2 it is 1, for r = 2 and n = 2 it is 2, for r = 1 and n = 3 it is 3, for r = 2 and n = 3 it is 4, for r = 3 and n = 3 it is 5, for r = 1 and n = 4 it is 6, etc. Unfortunately, some LAS files start numbering r and n at 0, only have return numbers r, or only have number of return of given pulse counts n. We therefore complete the table to also map invalid combinations to be used as a context. Furthermore with up to 15 returns there would be too many different combinations (namely 120). Therefore we map combinations of many returns pulses to the same number m and also complete the table for all combinations of r and n as shown below:
const U8 number_return_map_4bit[16][16] =
{
{ 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 },
{ 14, 0, 1, 3, 6, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10 },
{ 13, 1, 2, 4, 7, 11, 11, 11, 11, 11, 11, 10, 10, 10, 10, 10 },
{ 12, 3, 4, 5, 8, 12, 12, 12, 12, 11, 11, 11, 11, 11, 11, 11 },
{ 11, 6, 7, 8, 9, 13, 13, 12, 12, 12, 12, 11, 11, 11, 11, 11 },
{ 10, 10, 11, 12, 13, 14, 14, 13, 13, 12, 12, 12, 12, 12, 12, 12 },
{ 9, 10, 11, 12, 13, 14, 15, 14, 13, 13, 13, 12, 12, 12, 12, 12 },
{ 8, 10, 11, 12, 12, 13, 14, 15, 14, 13, 13, 13, 13, 12, 12, 12 },
{ 7, 10, 11, 12, 12, 13, 13, 14, 15, 14, 14, 13, 13, 13, 13, 13 },
{ 6, 10, 11, 11, 12, 12, 13, 13, 14, 15, 14, 14, 14, 13, 13, 13 },
{ 5, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 14, 14, 14, 13, 13 },
{ 4, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15, 14, 14, 14 },
{ 3, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15, 15, 14, 14 },
{ 2, 10, 10, 11, 11, 12, 12, 12, 13, 13, 14, 14, 15, 15, 15, 14 },
{ 1, 10, 10, 11, 11, 12, 12, 12, 13, 13, 13, 14, 14, 15, 15, 15 },
{ 0, 10, 10, 11, 11, 12, 12, 12, 13, 13, 13, 14, 14, 14, 15, 15 }
};
However. This turned out to give too many contexts resulting in many tables and hardly used prediction models. Therefore the above table was finally simplified to only map to 6 different contexts.
const U8 number_return_map_6ctx[16][16] =
{
{ 0, 1, 2, 3, 4, 5, 3, 4, 4, 5, 5, 5, 5, 5, 5, 5 },
{ 1, 0, 1, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3 },
{ 2, 1, 2, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, 3, 3, 3 },
{ 3, 3, 4, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4 },
{ 4, 3, 4, 4, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4 },
{ 5, 3, 4, 4, 4, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4 },
{ 3, 3, 4, 4, 4, 4, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4 },
{ 4, 3, 4, 4, 4, 4, 4, 5, 4, 4, 4, 4, 4, 4, 4, 4 },
{ 4, 3, 4, 4, 4, 4, 4, 4, 5, 4, 4, 4, 4, 4, 4, 4 },
{ 5, 3, 4, 4, 4, 4, 4, 4, 4, 5, 4, 4, 4, 4, 4, 4 },
{ 5, 3, 4, 4, 4, 4, 4, 4, 4, 4, 5, 4, 4, 4, 4, 4 },
{ 5, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 4, 4, 4 },
{ 5, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 4, 4 },
{ 5, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 4 },
{ 5, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5 },
{ 5, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5 }
};
The *return level l* specifies how many returns there have already been for a given pulse prior to this return. Given only valid combinations for the return number r and the number of returns of given pulse n we could compute it as l = n − r. But we again use a completed look-up table as shown below to map invalid combinations for r and l to different contexts.
const U8 number_return_level_4bit[16][16] =
{
{ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 14, 15 },
{ 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 14 },
{ 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12 },
{ 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 },
{ 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 },
{ 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 },
{ 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 },
{ 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8 },
{ 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7 },
{ 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6 },
{ 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5 },
{ 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4 },
{ 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3 },
{ 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2 },
{ 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, 1 },
{ 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 }
};
However. This turned out to give too many contexts resulting in many tables and hardly used prediction models. Therefore the above table was simplified to only map to 8 different contexts.
const U8 number_return_level_8ctx[16][16] =
{
{ 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7 },
{ 1, 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7 },
{ 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7 },
{ 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7 },
{ 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7 },
{ 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7 },
{ 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 7, 7 },
{ 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7, 7 },
{ 7, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6, 7 },
{ 7, 7, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5, 6 },
{ 7, 7, 7, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5 },
{ 7, 7, 7, 7, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3, 4 },
{ 7, 7, 7, 7, 7, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2, 3 },
{ 7, 7, 7, 7, 7, 7, 7, 6, 5, 4, 3, 2, 1, 0, 1, 2 },
{ 7, 7, 7, 7, 7, 7, 7, 7, 6, 5, 4, 3, 2, 1, 0, 1 },
{ 7, 7, 7, 7, 7, 7, 7, 7, 7, 6, 5, 4, 3, 2, 1, 0 }
};
We also create a simple return context 'cpr' with four values for the *current point* that is used after its return number and its number of returns were compressed. Its values are single = 3, first = 1, last = 2, and intermediate = 0 and it is computed:
I32 cpr = (r == 1 ? 2 : 0); // is first ?
cpr += (r >= n ? 1 : 0); // is last ?
We encode the X and Y coordinates with a second order predictor. We predict the differences dX and dY to the X and Y coordinates of the previous point from the same scanner channel with recently occuring differences. We predicts dX and dY between as the median of the five immediately preceding differences of points from the scanner channel and with the same return map m. The intuition behind this is, for example, that single returns are always from a different laser pulse than the previous point and therefore have a wider spacing in X and/or Y than the middle of three returns.
Compression of classification layer
-----------------------------------
We compress the classification as a symbol between 0 and 255 using 64 different context. This context is computed as the five lowest bits of the classification of the previous point (from the same scanner channel) multiplied by two plus one when the previous point (from the same scanner channel) was a single return.
Compression of flags layer
--------------------------
We compress the classification as a 6 bit number or a symbol between 0 and 63 using 64 different context. The highest bit is the 'edge_of_flight_line' flag. The second higest bit is the 'scan_direction' flag. The lowest four bits are the four classification flags in the order listed in the LAS 1.4 specification. The context is simply the 6 bit number that represented the flags of the previous point (from the same scanner channel).
Compression of intensity layer
------------------------------
We compress the intensity with the difference integer compressor using four contexts that is simply the 'cpr' defined earlier. We use one of 8 possible previous intensities as the prediction. Which one is used is decided by the 'cpr' multiplied by two plus one if the GPS time in respect to the previous point (from the same scanner channel) has changed. All eight possible previous intensities are initialized to the intensity of the first point (from the same scanner channel) and updated after the intensity was compressed.
Compression of scan angle layer
------------------------------
Only if the **scan angle is different** (as indicated by the 7 bits earlier) we compress it with the difference integer compressor using two contexts that is whether the GPS time stamp of the previous point (from the same scanner channel) has changed (1) or not (0). We use the scan angle of the previous point (from the same scanner channel) as the prediction.
Compression of user data layer
------------------------------
We compress the user data layer which is an 8 bit number as a symbol between 0 and 255 using 64 different contexts. The context is simply the user data of the previous point (from the same scanner channel) divided by 4.
Compression of point source ID layer
------------------------------
Only if the **point source ID is different** (as indicated by the 7 bits earlier) we compress
Compression of GPS time layer
-----------------------------
Only if the **GPS time is different** (as indicated by the 7 bits earlier) we compress it the same way it was done in the old LASzip coder. The GPS times of a single flight line stored in aquisition order are a monotonically increasing sequence of double-precision floating-point numbers where returns of the same pulse have the same GPS time and where subsequent pulses have a more or less constant spacing in time. LASzip treats the double-precision floating-point GPS times as signed 64 bit integers and predicts the deltas between them. We store up to four previously compressed GPS times with corresponding deltas to account for repeated jumps in GPS times that can arise when multiple flight paths are merged with fine spatial granularity. Nothing needs to be encoded when the GPS times are identical. Otherwise we distinguishes several cases that are entropy coded with 515 symbols depending on if the current GPS time is
+ 0–500 : predicted using the current delta times 0 to 500.
+ 501–510 : predicted using the current delta times -1 to -10.
+ 511 : starting a new context.
+ 512–514 : predicted with one of the other three contexts.
For the first two cases we subsequently difference code the delta prediction and the actual delta. We starts a new context when the delta overflows a 32 bit integer. For that it difference codes the 32 higher bits of the current GPS time and the current context and stores the lower 32 bits raw. Otherwise it switches to the specified context (where the delta will not overflows a 32 bit integer) and recurses. The current deltas stored with each context are updated to the actual delta when they were outside the predictable range more than 3 consecutive times (i.e. bigger than 500 times the current delta, smaller than -10 times the current delta, or zero).
Compression of RGB layer
------------------------------
We compress the RGB values the same way it was done in the old LASzip coder. LAS uses unsigned 16 bit integers for the R, G, and B
channel. Some files—incorrectly—populate only the lower 8 bits so that the upper 8 bits are zero. Other files correctly multiply 8-bit colors with 256 so that the lower 8 bits are zero. The LASzip compressor therefore compresses the upper and lower byte of each channel separately. First it entropy codes (a) 6 bits that specify which of the six bytes have changed in comparison to the previous point (from the same scanner channel) and (b) one bit that specifies whether those changes are the same in all three channels (i.e. grey colors) as one symbol between 0 and 127. For all bytes that have changed it then entropy codes the difference to the respective previous byte (of the point from the same scanner channel) modulo 256. If the 7 bit symbol indicates that the changes are identical in all three channels only the R channel is compressed. Otherwise the channels are encoded in the order R, G, and B. Differences encoded in earlier channels are used to predict differences in later channels as there tends to be a correlation in the intensity across channels. For example, if there was a byte difference in the low byte of the R channel that difference is added to low byte of the G channel which - clamped to a 0 to 255 range - becomes the value to which the difference of the current low byte is computed.
Compression of NIR layer
------------------------------
We compress the NIR value using the same technique as done for the RGB values. First it entropy codes 2 bits that specify which bytes have changed as one symbol. For all bytes that have changed it then entropy codes the difference to the respective previous byte (of the point from the same scanner channel) modulo 256.
Compression of WavePacket layer
------------------------------
We compress the WavePackets the same way it was done in the old LASzip coder. They only occur in point types 9 and 10. So far there has been very little real-world demand for compressing LAS files containing waveform data simply due to a lack of data stored in this format (i.e. there is no a dedicated format for full waveform LiDAR called 'PulseWaves'). LASzip simply entropy codes the wave packet descriptor index, an unsigned byte that is zero if a point has no waveform and indexes the variable length record describing the format of the waveform otherwise. To compress the bytes offset to waveform data it entropy encodes one of 4 possible cases:
1) same as last offset
2) use last offset plus last packet size
3) difference to last offset is less than 32 bits
4) difference to last offset is more than 32 bits
In the first two cases no other information is needed. For the other two cases LASzip difference codes the 32 or the 64 bit numbers. The LASzip compressor difference coded all remaining fields. Only waveform packet size in bytes is an integer number. The return point waveform location, x(t), y(t), and z(t) are single-precision floating-point numbers whose 32 bits are treated as if they were a 32-bit integer.
Compression of "Extra Bytes" layer(s)
------------------------------
We compress the "Extra Bytes" the same way it was done in the old LASzip coder but each byte is compressed into its own layer. A LAS point has “Extra Bytes” when the LAS header specifies a point size larger than required by the respective point type. Each “Extra Byte” is entropy encoded with its own context as the difference to the “Extra Byte” from the previous point (from the same scanner channel) modulo 256. Treating them as individual bytes is the best that the LASzip compressor can do as there is not always a description what these “Extra Bytes” may mean. Six “extra bytes”, for example, could be a single-precision float storing the echo width followed by an unsigned short storing the normalized reflectivity. Or it could be an unsigned short storing a tile index followed by an unsigned integer storing the original index of the point.