Comparing classification accuracy of NDVI with DEM derived attributes using multi-scalar approach in Geographic Information Systems

Increase in resolution of Digital Elevation Models (DEMs), multispectral data generated by remote sensing satellites and Lidar have caused a rapid increase in the amount of data that can be processed by GIS software. However, the GIS tools still use a fixed 3-by-3 window to evaluate attributes such as slope, elevation and curvature. These tools do not scale well to a 5m DEM resolution compared to 30m resolution used earlier. Results generated usually contain a lot of noise. To overcome this problem researchers prefer to resample DEMs to a lower resolution which in turn decreases their accuracy. In this paper we test the efficiency of sliding window-based aggregation technique on DEM derived attributes on a multi-scalar level. We use an iterative aggregation algorithm to obtain these attributes for multiple window sizes 4, 8 and 16. Results obtained from previous iteration are reused in the next iteration. Hence, we achieve logarithmic efficiency compared to the existing resampling technique in ArcMap. Results are compared with resampled DEM data at window sizes 3, 9 and 15. Random Forest and Maximum Likelihood classification have been applied to all the outputs with Normalized Difference Vegetation Index (NDVI) as the dependent variable. Classification on results from the proposed method achieves a higher accuracy compared to traditional resampling technique. Window size 4 produces the highest accuracy of 74.47% using the Random Forest based classification compared to 46.31% generated using output from ArcMap. The accuracy remains fairly consistent across all the window sizes for the proposed method while the classification accuracy generated from resampled data falls drastically with higher window size. This shows that sliding window based aggregation is both scalable and accurate.