Canadian Forest Service Publications
Quantifying the contribution of spectral metrics derived from digital aerial photogrammetry to area-based models of forest inventory attributes. 2019. Tompalski, P., White, J.C., Coops, N.C., Wulder, M.A. Remote Sensing of Environment 234, 111434.
Year: 2019
Issued by: Pacific Forestry Centre
Catalog ID: 40012
Language: English
Availability: PDF (download)
Available from the Journal's Web site. †
DOI: 10.1016/j.rse.2019.111434
† This site may require a fee
Abstract
Digital aerial photogrammetry (DAP) has demonstrated utility across a range of forest environments as an alternative data source to airborne laser scanning (ALS) for estimating forest inventory attributes in an area-based approach. In this context, metrics are typically derived from the DAP point cloud in a manner analogous to that of ALS data. However, image matching algorithms also allow for spectral information from the image data to be transferred to the point cloud. Herein, we quantify the contribution of this spectral information to the area-based prediction of five forest inventory attributes: Lorey's mean height, quadratic mean diameter, basal area, gross volume per ha, and stems per ha in a highly productive coastal temperate rainforest on Vancouver Island, British Columbia, Canada. Using ground plots and ALS-derived area-based estimates as reference, we compare plot-level predictions generated using (i) DAP point cloud metrics, (ii) DAP spectral metrics, and (iii) combinations of DAP point cloud and spectral metrics. In addition to prediction accuracy, we assessed variable importance to identify those metrics that were most informative for the developed models. Our results indicated that for models generated using DAP data, prediction accuracy was greatest when the point cloud-based metrics were incorporated. Models that incorporated both point cloud and spectral information were only slightly more accurate than models based on point cloud metrics only. We found that the improvement in accuracy was not observed for all stand attributes. The highest increase in accuracy for models combining point cloud and spectral metrics was observed for quadratic mean diameter, basal area per hectare, and stem volume per hectare, with change in relative root mean square error of −1.3%, −1.75%, and −0.23%, respectively. Models derived with spectral metrics only had the lowest accuracy with R2 values never exceeding 0.25. Analysis of the variable importance indicated that point cloud metrics are markedly more important than spectral metrics. We conclude that the benefit of the additional spectral information in this forest environment is negligible, and the effort to derive the spectral information cannot be justified for operational applications. Our results confirm those of other studies in other environments that have likewise found minimal benefit to the incorporation of DAP spectral information in area-based estimation.
Plain Language Summary
Digital aerial photogrammetry (DAP) has been demonstrated as an alternative data source to airborne laser scanning (ALS) for forest inventory. However, image matching algorithms also allow for spectral information from the image data to be transferred to the point cloud. By quantifying the contribution of the spectral information to the area-based prediction of five forest inventory attributes we found that models that incorporated both point cloud and spectral information were only slightly more accurate than models based on point cloud metrics only, with decrease in relative root mean square error of less than 2%. Analysis of the variable importance indicated that point cloud metrics are markedly more important than spectral metrics. We conclude that the benefit of the additional spectral information in this forest environment is negligible when predicting forest structure attributes such as height and volume, and the effort to derive the spectral information cannot be justified for operational applications. In short: • Area-based models generated for five forest attributes • Attributes were modelled using different sets of point cloud and spectral metrics • Spectral metrics provided no significant increase in model accuracy • Point cloud metrics were markedly more important than spectral metrics
