Developers Corner: Improving GeoTools/GeoServer raster reprojection performance
Ciao a tutti,
we hope everybody had nice seasonal holidays and is back fresh and ready to move on with another year of activity.
Raster reprojection is a quite heavy process in which every single pixel of the original image has to be mapped into a new position in the target raster. Here is a visual representation of a small set of pixels, before and after the reprojection (in this case, from WGS84 to polar stereographic):
Now, if you think about it, the transformation applied to one pixel and the one applied to the next one are basically the same, the difference is going to be minimal. If we only had a small bunch of pixels we could compute a simple translation and apply it to them all.
Having more pixels we could try to approximate the overall transformation with a simple linear one, which in two dimensional space is known as an affine transformation.
If the area gets larger a single affine transformation would not be good enough anymore, but we could use many of them to fit different areas of the raster.
Have a look at the following one dimensional representation to get the gist of the idea (this is known as piecewise linear function):
The trick to apply this approach if finding out where and how much to subdivide the original area into discrete pieces that can use a single affine transform: to do so we devised a simple error estimation method that proved to work well with all common map projections.
So we start by evaluating the linear transformation error over the entire raster area, if it’s larger than a certain target we split the area and evaluate again, and recurse this way until the approximation error has been put under control.
The optimized method proved to be quite effective, drastically reducing the time it takes to reproject a raster. We made some tests with GeoServer using the same data and the same reprojection test used in the FOSS4G 2010 WMS shootout measuring a six times speedup:
We also compared the images generated by the original and optimized method finding there was hardly any human noticeable difference, and very little difference that could be machine detected at all.
For all of you that want to dig into the nitty-gritty details, we have a full tech report that you can inspect.
There is still plenty that can be done to optimize further raster data reading and in memory transformations to reach even higher performance and scalability. Interested? Let us know!
The GeoSolutions team