@article{fdi:010091201,
  title = {{U}noccupied aerial system ({UAS}) {S}tructure-from-{M}otion canopy fuel parameters : multisite area-based modelling across forests in {C}alifornia, usa},
  author = {{R}eilly, {S}. and {C}lark, {M}. {L}. and {L}oechler, {L}. and {S}pillane, {J}. and {K}ozanitas, {M}. and {K}rause, {P}. and {A}ckerly, {D}. and {B}entley, {L}. {P}. and {O}liveras {M}enor, {I}mma},
  editor = {},
  language = {{ENG}},
  abstract = {{T}here is a pressing need for well-informed management to reduce wildfire hazard and restore fire's beneficial ecological role in the {M}editerranean- and temperate-climate forests of {C}alifornia, {USA}. {T}hese efforts rely upon the accessibility of high spatial and temporal resolution data on biomass and canopy fuel parameters such as canopy base height ({CBH}), mean canopy height, canopy bulk density ({CBD}), canopy cover, and leaf area index ({LAI}). {R}emote sensing using unoccupied aerial system {S}tructure-from-{M}otion ({UAS}-{S}f{M}) presents a promising technology for this application due to its accessibility, relatively low cost, and possibility for high temporal cadence. {H}owever, to date, this method has not been studied in the complex mosaic of forest types found across {C}alifornia. {I}n this study we examined the capacity of structural and multispectral information obtained from {UAS}-{S}f{M}, in conjunction with machine learning methods, to model aboveground biomass and forest canopy fuel structural parameters using an area-based approach across multiple sites representing a diversity of forest types in {C}alifornia. {B}ased on correlations with field measurements, fuel parameters separated into vertical (biomass, {CBH}, and mean height) and horizontal ({LAI}, {CBD}, canopy cover) groups. {UAS}-{S}f{M} random forest models performed well for modelling the vertical structure canopy fuels parameters ({R}2 0.69-0.75). {T}hese models exhibited strong performance in comparison to {ALS}, as well as when transferred to a novel site. {V}ertical structure predictors were prominent in these models, and did not improve with the addition of spectral predictors. {UAS}-{S}f{M} random forest models of horizontal structure parameters mainly used raster-based spectral indices (primarily {NDVI}) and had relatively low performance ({R}2 0.49-0.59). {I}n addition, these models underperformed {ALS} and had poor performance when applied to a novel site. {W}hen applied to a region with widespread {UAS}-{S}f{M} coverage, models from both groups successfully produced contiguous maps that could be used for modelling fire behavior or in management decision making and monitoring. {T}hese findings indicate that {UAS}-{S}f{M}, without the need for multispectral sensors, is well suited for mapping area-based vertical-structure canopy parameters across diverse landscapes supporting a wide range of forest types. {I}n contrast, the identification of spectral mean variables for modelling horizontal structure canopy fuels suggests the potential of multi- or hyperspectral sensors or high-resolution satellite imagery for meeting management information needs.},
  keywords = {{UAS} ; {S}f{M} ; {M}achine learning ; {C}anopy fuels ; {B}iomass ; {C}alifornia forests ; {ETATS} {UNIS} ; {CALIFORNIE}},
  booktitle = {},
  journal = {{R}emote {S}ensing of {E}nvironment},
  volume = {312},
  numero = {},
  pages = {114310 [16 ]},
  ISSN = {0034-4257},
  year = {2024},
  DOI = {10.1016/j.rse.2024.114310},
  URL = {https://www.documentation.ird.fr/hor/fdi:010091201},
}