@article{fdi:010063635,
  title = {{C}anopy height estimation in {F}rench {G}uiana with {L}i{DAR} {ICES}at/{GLAS} data using principal component analysis and random forest regressions},
  author = {{F}ayad, {I}. and {B}aghdadi, {N}. and {B}ailly, {J}. {S}. and {B}arbier, {N}icolas and {G}ond, {V}. and {E}l {H}ajj, {M}. and {F}abre, {F}. and {B}ourgine, {B}.},
  editor = {},
  language = {{ENG}},
  abstract = {{E}stimating forest canopy height from large-footprint satellite {L}i{DAR} waveforms is challenging given the complex interaction between {L}i{DAR} waveforms, terrain, and vegetation, especially in dense tropical and equatorial forests. {I}n this study, canopy height in {F}rench {G}uiana was estimated using multiple linear regression models and the {R}andom {F}orest technique ({RF}). {T}his analysis was either based on {L}i{DAR} waveform metrics extracted from the {GLAS} ({G}eoscience {L}aser {A}ltimeter {S}ystem) spaceborne {L}i{DAR} data and terrain information derived from the {SRTM} ({S}huttle {R}adar {T}opography {M}ission) {DEM} ({D}igital {E}levation {M}odel) or on {P}rincipal {C}omponent {A}nalysis ({PCA}) of {GLAS} waveforms. {R}esults show that the best statistical model for estimating forest height based on waveform metrics and digital elevation data is a linear regression of waveform extent, trailing edge extent, and terrain index ({RMSE} of 3.7 m). {F}or the {PCA} based models, better canopy height estimation results were observed using a regression model that incorporated both the first 13 principal components ({PC}s) and the waveform extent ({RMSE} = 3.8 m). {R}andom {F}orest regressions revealed that the best configuration for canopy height estimation used all the following metrics: waveform extent, leading edge, trailing edge, and terrain index ({RMSE} = 3.4 m). {W}aveform extent was the variable that best explained canopy height, with an importance factor almost three times higher than those for the other three metrics (leading edge, trailing edge, and terrain index). {F}urthermore, the {R}andom {F}orest regression incorporating the first 13 {PC}s and the waveform extent had a slightly-improved canopy height estimation in comparison to the linear model, with an {RMSE} of 3.6 m. {I}n conclusion, multiple linear regressions and {RF} regressions provided canopy height estimations with similar precision using either {L}i{DAR} metrics or {PC}s. {H}owever, a regression model (linear regression or {RF}) based on the {PCA} of waveform samples with waveform extent information is an interesting alternative for canopy height estimation as it does not require several metrics that are difficult to derive from {GLAS} waveforms in dense forests, such as those in {F}rench {G}uiana.},
  keywords = {{L}i{DAR} ; {ICES}at/{GLAS} ; canopy height ; tropical forest ; {F}rench {G}uiana ; {GUYANE} {FRANCAISE}},
  booktitle = {},
  journal = {{R}emote {S}ensing},
  volume = {6},
  numero = {12},
  pages = {11883--11914},
  ISSN = {2072-4292},
  year = {2014},
  DOI = {10.3390/rs61211883},
  URL = {https://www.documentation.ird.fr/hor/fdi:010063635},
}