 This presentation will give you some pointers on projecting raster data. That is, when we want to take a raster data set and project it from one coordinate system to another. Prior to projecting your raster data, you should have a really good understanding about raster data types, specifically if you're working with thematic or discrete raster data or continuous raster data set. Moving over to ArcGIS Pro, I've got two different data sets I'm going to use for the demo. One is a thematic raster data set, a seven-category land cover classification, and the other is a high-resolution digital elevation model, representing the topographic surface that was derived from LIDAR data. Projecting raster data is different than projecting vector data because we're creating a new raster in which the cells are in a different location. To project your raster, you have to use the Project Raster tool. This is different than the Project tool that you use for vector data. Starting with the DEM, I'm going to transform the coordinate system for both these raster data sets from state plane to UTM. Because we can't stretch the data like we can with vector data when going from one coordinate system to another, I have to choose a resampling method. For both these data sets, I'm going to show you the effect of a nearest neighbor resampling and a cubic convolution resampling approach. ArcGIS Pro currently offers four different ways of resampling your data. Nearest neighbor, bilinear interpolation, cubic convolution, and majority resampling. The rule of thumb is to use nearest neighbor or majority resampling for thematic or discrete data and cubic convolution for continuous data, should almost never need to use bilinear interpolation. So in this first run of the DEM, I'm creating a new DEM in the WGS84 UTM Zone 16 coordinate system. Using nearest neighbor resampling, and because I'm going from a 1.5 foot data set to a meter data set, you can see the cell size has changed to reflect the 1.5 equivalent in meters. For comparison purposes, I'm also running this using the cubic convolution resampling technique. The only parameter altered is the resampling technique. In nearest neighbor resampling, the cell in the output raster closest in location to the input raster gets the value, as opposed to cubic convolution, which takes a weighted sampling of the 16 surrounding cells. Now we're going to repeat the same process for our land cover data set. Once again, we're projecting it into the UTM Zone 16 WGS84 coordinate system using both nearest neighbor and then cubic convolution as the resampling technique. And then we'll have a look at the output to see how the resampling technique influences the quality of the output raster. To examine the impact on the nearest neighbor resampling, I'm going to zoom into the land cover data and swipe back and forth between the original land cover data set and the projected raster data set using nearest neighbor resampling. Overall, the two data sets look rather similar. Now keep in mind they are different colors, but that's just due to the symbology. However, you will notice that because we created a new raster data set as part of the projection process and the coordinate system is different, that the cells are in slightly different locations and in some instances features may appear to have been shifted. Generally you can project vector data back and forth without issues, but raster data is fundamentally altered when you run the project tool. Now let's take a look at what happens when we compare the original land cover to the cubic convolution output. You're going to notice that we have all these strange, individually colored pixels next to the roads, trees, and other boundary features. As I mentioned earlier, cubic convolution takes a weighted average of the surrounding cell values. Because this is discrete arithmetic data in which each cell represents an individual land cover code, a weighted average makes no sense, because now I've fundamentally altered not just the position of the land cover feature, but I've changed the type of land cover feature that it is. Hopefully you now realize that cubic convolution resampling should never be used for discrete arithmetic data because it alters the cell values and will corrupt your data. Moving over to the DEMs, the first thing you'll notice is there are slightly different ranges between the three data sets. Once again, that's because we've altered the pixel values. Swiping back and forth between the original DEM and the new DEM that's projected using nearest neighbor resampling shows a few, but it seems to be relatively minor differences. Switching over to the cubic convolution data set and swiping back and forth between the original and the projected raster, we also don't see any differences that appear to be substantial. When examining the effects of the resampling technique on topographic data, the use of a hillshade can help reveal imperfections that would otherwise go unnoticed. Here I'm running the hillshade function on both the nearest neighbor projected raster and the hillshade function on the cubic convolution projected raster, creating two new hillshade layers. A tip on working with hillshades, just make sure you don't have the zero value set to transparent. As a result, the zero values in your hillshade will show through to the underlying layers. In examining the hillshade data, we see that for the nearest neighbor resampling approach, we have a very undesirable checkerboard pattern. The nearest neighbor resampling technique does a great job of preserving the original pixel values, but it distorts the spatial representation when we're using it to project data from one coordinate system to another. Cubic convolution, although it alters the values, produces a much more desirable representation. If you are working with a land cover data set that is thematic or contains discrete coded values like land cover data, use nearest neighbor or majority resampling when projecting your raster. If you're working with a continuous raster data set, something like elevation, then you should use cubic convolution resampling. Bilinear interpolation is rarely used because it was designed for a time when computers did not have the processing power to use the cubic convolution resampling method.