I have a mosaic of a 1m LiDar derived DEM. I need to output a subset of the data as a 10m DEM. I am currently using the aggregate tool in ARCGIS 10 to produce a mean value for each new 10m pixel. Any advice on whether this is the best technique for such as task? Is the mean value the best approach with this type of data?
A common mistake (that I've made too) is to down-sample a raster using the resample tool with bilinear interpolation. See this answer for an explanation why this is not good. A raster can be down-sampled in three steps.
The first step might not be required. Reproject the raster to the target extents. Use bilinear interpolation, and keep the output cell size the same as the input resolution (e.g., 1 m). Use the registration point to "snap" the raster corners to the projection. The output extents can be specified in the "Environments", and I suggest specifying the extents with a multiple of 10 m (or whatever resolution). These extents will control where the statistics are determined for the final raster.
Perform Block Statistics (found in Spatial Analyst Tools > Neighborhood). Use a rectangle with 10 cells for both height and width, and choose "MEAN" for a statistic type. Try different shapes and types, if you want. The cell size is the down-sampling ratio.
Since block statistics does not change the raster resolution, the last step is to Resample (found in Data Management Tools > Raster > Raster Processing). Choose 10 m, and use "NEAREST" to pick the block statistic in the centre of the cell.
Steps 2 and 3 can be replaced with's Curtvprice's suggestion to use the Aggregate tool, which will effectively yield the same results using rectangle means.
You can cut a step out of the process by using the Aggregate tool, which avoids having to replicate all those summary value cells, and the Resample.
All the points mentioned above are important to take note of and I completely agree that bilinear resampling is quite problematic. Although, I am curious why nobody is discussing cubic convolution? The problem with using a block function is that the mean is quite irrelevant when the distribution is non-normal or multimodal, as expected with a lidar derived DEM.
If you have access to the original lidar data, just interpolate the data to the desired resolution using the "Topo to Raster" tool in ArcGIS. If you only have access to the 1m DEM raster, it seems like the best, albeit least efficient, method would be to convert the raster to points and use a thin plate or Bi-cubic spline. This would allow for the resample neighborhood to fit a nonlinear curve to the data.
Alternatively, you could smooth the 1m raster using a Gaussian kernel, approximating the size of your desired resample resolution (10x10), and then a bilinear resample would be much more appropriate. This approach would allow you have direct control over the smoothing parameter and would result in a "locally" normal distribution where the mean becomes relevant as an indicator of central tendency and a linear fit is supported.