Title | Fusion of LiDAR elevation and canopy derivatives with other imagery sources for wetland classification using Random Forest classification:
A Case Study of the Mer Bleue Wetland, Ottawa, Ontario |
| |
Author | Millard, K; Richardson, M |
Source | 33rd Canadian Symposium on Remote Sensing, abstracts; by Canadian Symposium on Remote Sensing; 2012 p. 30 Open Access |
Links | Online - En ligne
|
Links | Abstracts (PDF, 1.22 MB)
|
Year | 2012 |
Alt Series | Earth Sciences Sector, Contribution Series 20140079 |
Meeting | 33rd Canadian Symposium on Remote Sensing; Ottawa; CA; June 11-14, 2012 |
Document | book |
Lang. | English |
Media | on-line; digital |
File format | pdf |
Province | Ontario |
NTS | 31G/05 |
Area | Ottawa; Mer Bleue |
Lat/Long WENS | -76.0000 -75.5000 45.5000 45.2500 |
Subjects | geophysics; remote sensing; satellite imagery; wetlands; vegetation |
Released | 2012 01 01 |
Abstract | LiDAR is best known for its use in the creation of high resolution digital elevation models (DEMs) (Toyra et al, 2003) however various derivatives can also be calculated from LiDAR DEMs (e.g. measures
of relative topographic position) that can help us characterize the micro-topographic features of a wetland surface (Richardson et al, 2010). Discrete return LiDAR can provide information about the canopy structure through the collection and
classification of multiple laser pulse returns (Hopkinson et al, 2006) and the intensity of the return has been shown to be useful in the determination of relative soil moisture and vegetation characteristics (Korpela, 2009). In this study, the
combination of LiDAR DEM and canopy model derivatives in conjunction with other types of remotely sensed data (optical and radar) are considered in the classification of the Mer Bleue wetland, Ottawa, Ontario. While this multi-source dataset allows
extraction of unique information from each type of data and derivative, classification of multi-source data can be challenging with traditional classification methods (Gislason et al, 2006).
With a focus on distinguishing wetland classes as
well as the ability to separate wetland from upland, we use Random Forest (Breiman, 2001), an ensemble classifier that creates multiple decision trees to classify our multi-source dataset. Variable importance plots resulting from the Random Forest
classifier help us determine which input variables are most appropriate and useful for classifying the different wetland and upland classes and help interpret the physical attributes of the wetland surface, which can subsequently be linked to the
unique geomorphic and ecohydrological characteristics of different landcover types within this hetereogeneous peatland complex. |
GEOSCAN ID | 294588 |
|
|