Lab 9 - Accuracy of DEMs

This week we revisited the topic of accuracy, which we touched on in Lab 1 for horizontal accuracy. This time we were looking at vertical accuracy of LIDAR based and interpolated DEMs. According to the National Digital Elevation Program (NDEP) guidelines, we can assess the quality of a DEM by using a specified number of checkpoints of higher accuracy and comparing those values to the values extracted from the DEM. Error can be measured using Root Mean Square Error (RMSE), and accuracy and bias can be further determined using percentiles and mean difference results.

For the first part of the lab, I used a provided LIDAR data DEM to extract values to the location of field data points. The field data was collected in a cluster sample pattern, with points collected fairly evenly between land cover types of a) bare earth and low grass, b) high grass, weeds and crops, c) low trees and shrubs, d) fully forested, and e) urban, with a extra points collected in the fully forested category. Using the absolute difference between the two datasets, I calculated the RMSE, 68th percentile, and 95th percentile for each land type. Results showed a slightly higher RMSE for the fully forested and urban land cover types. Next I used the mean difference from the original data to look at whether there was any bias in the data. I also displayed the difference values with a diverging color ramp to visually look for at any bias patterns. In both bases, statistics and visuals, there was a higher proportion of urban and fully forested land type classes that were negative values, suggesting along with the earlier results that there was bias and lowered accuracy with respect to fully forested and urban land cover types. These results are as expected based on the well-studied relationship between increased vegetation and feature density causing a decrease in LIDAR laser capabilities to accurately assess ground elevation.

The second part of the lab consisted of creating DEMs from provided point data using three different interpolation methods: IDW, Spline, and Kriging. Using a 5% sample point that was held out of the initial interpolated points for accuracy assessment, I calculated the average difference and average absolute differences to get a general idea of which interpolation method did a better job predicting elevation. Based on those calculations and results, all three interpolations show a negative average difference value (-1.651 for IDW, -2.774 for Spline, -2.076 for Kriging), suggesting they are all overestimating elevation. For absolute difference the largest difference was Kriging at 10.142, then IDW at 9.840, then Spline at 8.256. Considering both calculations together, I think IDW does a better job predicting elevation and the 5% sample points than the other two methods, as it has the smallest overestimation amount and a middle absolute difference value.

Screenshot of IDW interpolated DEM and 5% withheld sample points used to assess quality of DEM.