An Assessment of Aerial Survey Acquisition Methods for Generating Forest Inventory Models

By: David Sandilands - Master of Science in Forest Resources
Advisory Committee:
Daniel Hayes - Assistant Professor of Geospatial Analysis & Remote Sensing; Director, Barbara Wheatland Geospatial Analysis Laboratory
Aaron Weiskittel - Professor of Forest Biometrics and Modeling; Director, Center for Research on Sustainable Forests
Erin Simons-Legaard - Assistant Research Professor in Forest Landscape Modeling

08/31/2011 - 12/15/2017

3D point cloud colored by RGB imagery

Structure from motion (SfM) derived three dimensional (3D) point clouds have performed well in comparison to airborne laser scanning (ALS) in generating timber inventory models, yet little has been done to assess the viability of a wide range of aerial image acquisition methods for developing these models. Remote sensing factors such as sensor reflectance characteristics, perspective, overlap, and resolution as well as software processing parameters have cost implications both in terms of time and equipment investment. The remote sensing factors analyzed in this study included R/G/B vs. NIR/G/B imagery, vertical vs. oblique perspectives, 80%/65% vs. 60%/30% overlap, and ground sample distance (GSD) levels of 6cm, 12cm, and 24cm. The effect of point cloud density was also assessed. Ancillary factors such as slope, aspect, site quality, forest composition, and spatial patterns relating to georectification accuracy may also play roles in influencing inventory models and their potential influences were explored.

Research was conducted at the 40 hectare Holt Research Forest (HRF) located in Arrowsic, Maine which is characterized as representing oak-pine and coastal spruce-fir forest ecosystems. Chapter 1 analysis was performed on 10m X 10m cells across the study area and twenty-four SfM datasets were evaluated to assess their ability to estimate canopy height models (CHMs) based on maximum height and average height. Correlation values ranged from 0.72 to 0.98 between SfM-derived CHMs and those generated by ALS. Bias plots comparing SfM to ALS canopy height models indicated linear relationships for most comparisons across the range of height values and SfM datasets tended to over-estimate average height. Ancillary factors were shown to have statistically significant effects on the comparisons between SfM and ALS, though the extents to which are unknown and would require more in-depth analysis to determine. Comparisons were also affected by spatial patterns due to georectification errors which were more prevalent in vertical datasets composed of high image overlap.

The second half of this study evaluated SfM-derived area-based predictive models of gross timber volume in comparison to ALS-derived models and observed volume calculated from stem-mapped field plots. In addition to the remote sensing and ancillary factors examined in Chapter 1, the effects of three cell size categories (10m X 10m, 20m X 20m, and 30m X 30m), percent softwood composition, and stem density were analyzed. Volume was predicted by SfM with accuracies comparable to those produced by ALS irrespective of acquisition method and cell size. Volume prediction model RMSE averaged 66.1% for 10m by 10m cells, 40.8% for 20m X 20m cells, and 32.6% for 30m X 30m cells. Ancillary factor analysis indicated that percent softwood cover and trees per hectare (TPH) were highly influential on the amount of error between SfM predicted and observed volumes. As both TPH and percent softwood composition increased, SfM models were increasingly likely to trend from under to over-prediction. Additional work should be conducted over a variety of forest type and landscape conditions to better characterize trends.


Dave Sandilands