Journal
REMOTE SENSING OF ENVIRONMENT
Volume 209, Issue -, Pages 497-509Publisher
ELSEVIER SCIENCE INC
DOI: 10.1016/j.rse.2018.02.070
Keywords
Airborne Snow Observatory; Waveform lidar; Drylands; Plant functional types; Random forest; Machine learning; Feature selection; Ensemble; Carbon
Funding
- NASA Terrestrial Ecology [NNX14AD81G]
- NASA Earth and Space Science Fellowship [17-EARTH17F-0209]
- Division Of Environmental Biology [1550916] Funding Source: National Science Foundation
Ask authors/readers for more resources
Accurate classification of plant functional types (PFTs) reduces the uncertainty in global biomass and carbon estimates. Airborne small-footprint waveform lidar data are increasingly used for vegetation classification and above-ground carbon estimates at a range of spatial scales in woody or homogeneous grass and savanna ecosystems. However, a gap remains in understanding how waveform features represent and ultimately can be used to constrain the PFTs in heterogeneous semi-arid ecosystems. This study evaluates lidar waveform features and classification performance of six major PFTs, including shrubs and trees, along with bare ground in the Reynolds Creek Experimental Watershed, Idaho, USA. Waveform lidar data were obtained with the NASA Airborne Snow Observatory (ASO). From these data we derived waveform features at two spatial scales (1 m and 10 m rasters) by applying a Gaussian decomposition and a frequency-domain deconvolution. An ensemble random forest algorithm was used to assess classification performance and to select the most important waveform features. Classification models developed with the 10 m waveform features outperformed those at 1 m (Kappa (kappa) = 0.81-0.86 vs. 0.60-0.70, respectively). At 1 m resolution, lidar height features improved the PFT classification accuracy by 10% compared to the analysis without these features. However, at 10 m resolution, the inclusion of lidar derived heights with other waveform features decreased the PFT classification performance by 4%. Pulse width, rise time, percent energy, differential target cross section, and radiometrically calibrated backscatter coefficient were the most important waveform features at both spatial scales. A significant finding is that bare ground was clearly differentiated from shrubs using pulse width. Though the overall accuracy ranges between 0.72 and 0.89 across spatial scales, the two shrub PFTs showed 0.45-0.87 individual classification success at 1 m, while bare ground and tree PFTs showed high (0.72-1.0) classification accuracy at 10 m. We conclude that small-footprint waveform features can be used to characterize the heterogeneous vegetation in this and similar semi-arid ecosystems at high spatial resolution. Furthermore, waveform features such as pulse width can be used to constrain the uncertainty of terrain modeling in environments where vegetation and bare ground lidar returns are close in time and space. The dependency on spatial resolution plays a critical role in classification performance in tree-shrub co-dominant ecosystems.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available