Performance of non-parametric algorithms for spatial mapping of tropical forest structure

• Published in: Carbon balance and management 2016-08, Vol.11 (1), p.18-18, Article 18
• Main Authors: Xu, Liang, Saatchi, Sassan S., Yang, Yan, Yu, Yifan, White, Lee
• Format: Article
• Language: English
• Subjects: Canopy height; Earth and Environmental Science; Ecosystems; Environment; Environmental Management; Forestry; Gabon; Lidar; Spatial mapping; Tropical forests
• ISSN: 1750-0680; 1750-0680
• DOI: 10.1186/s13021-016-0062-9

Background Mapping tropical forest structure is a critical requirement for accurate estimation of emissions and removals from land use activities. With the availability of a wide range of remote sensing imagery of vegetation characteristics from space, development of finer resolution and more accurate maps has advanced in recent years. However, the mapping accuracy relies heavily on the quality of input layers, the algorithm chosen, and the size and quality of inventory samples for calibration and validation. Results By using airborne lidar data as the “truth” and focusing on the mean canopy height (MCH) as a key structural parameter, we test two commonly-used non-parametric techniques of maximum entropy (ME) and random forest (RF) for developing maps over a study site in Central Gabon. Results of mapping show that both approaches have improved accuracy with more input layers in mapping canopy height at 100 m (1-ha) pixels. The bias-corrected spatial models further improve estimates for small and large trees across the tails of height distributions with a trade-off in increasing overall mean squared error that can be readily compensated by increasing the sample size. Conclusions A significant improvement in tropical forest mapping can be achieved by weighting the number of inventory samples against the choice of image layers and the non-parametric algorithms. Without future satellite observations with better sensitivity to forest biomass, the maps based on existing data will remain slightly biased towards the mean of the distribution and under and over estimating the upper and lower tails of the distribution.