Remote Sensing based prediction of forest attributes

Loïc Dutrieux, 2015-10-20

ROBIN science meeting, Brussels

Julian Equihua, Lammert Kooistra, Martin Herold, Mariso Toledo, Nataly Ascarunz, Gerbert Roerink, Marielos Peña-Claros, Lourens Poorter

Why do we need maps of forest attributes?

Is that sufficient to:

  • Find spatial patterns ?
  • Design region specific policy actions ?
  • Limitations

  • Lacks spatial details
  • Spatially discrete
  • Costly to acquire

Asner, Gregory P., et al. "High-resolution forest carbon stocks and emissions in the Amazon." Proceedings of the National Academy of Sciences 107.38 (2010): 16738-16742.

The Remote Sensing scaling recipe

Take

  • 220 in-situ forest sites
  • 28 remote sensing preditors
  • A random forest model
    Step by step:

  • Select study site
  • Select and prepare remote sensing layers
  • Extract remote sensing variables for every in-situ location
  • Train random forest model
  • Select most important predictors
  • Predict

Study site

In-situ variables

(The variables we'll try to predict)

Variable Category
Species composition (First two DCA* axes) Floristic
Species richness Floristic
Emergent density Functional group
Liana infestation Functional group
Palm density Functional group
Mean DBH Structure
Tree density Structure
Basal area Structure

*Detrended Correspondence Analysis (DCA) summarizes species count for every site into a score.

Toledo, 2010. Neotropical lowland forests along environmental gradients

Remote sensing predictors

A set of ecologically relevant variables

  • Moderate resolution datasets (1km²)
  • A lot of them derived from MODIS
  • Available globally

Remote sensing predictors

Variable Expected contribution Number of predictors
Surface reflectance Leaf chemistry 14
LAI Foliage structure 1
Tree height Vertical structure 1
NPP Productivity 1
Phenology (HANTS) Intra-annual dynamics 5
Terrain metrics Topographic conditions 4
Cannopy clumping Canopy structure 1
Tree cover Forest density 1

Data extraction

Data extraction - considerations of scale

  • We assume sub-pixel spatial homogeneity
  • Spatial aggregation required due to clustering of in-situ sites
  • 104 calibration points after aggregation

Random forest ?










What is random forest?

  • An ensemble learning method
  • Builds multiple regression trees
  • Good prediction performances
  • Relatively interpretable (variable importance)
  • Quite insensitive to potentially irrelevant predictors
  • Well suited to such exploratory data mining approaches
  • It's pure coincidence we're using it to predict forest attributes

Results

Modelling performances assessment

Summary of variables remote sensing potential

Variable Category
Species composition (First two DCA* axes) Floristic
Species richness Floristic
Emergent density Functional group
Liana infestation Functional group
Palm density Functional group
Mean DBH Structure
Tree density Structure
Basal area Structure

What now?

We have spatially continuous information on floristic properties at 1km resolution, what can we do with it?

Does species diversity matter for biomass?



According to this correlation map:

It depends where !



Avitabile V., Herold M., Heuvelink G., Lewis S.L., Phillips O.L., Asner G. P. et al., 2015. An integrated pan-tropical biomass maps using multiple reference datasets. Global Change Biology (in review)

Metzger, Marc J., et al. "A high‐resolution bioclimate map of the world: a unifying framework for global biodiversity research and monitoring." Global Ecology and Biogeography

Wrapping up - Conclusions - What have we learnt?

  • We can predict biodiversity using remote sensing (Using a relatively simple approach)
  • These maps can be used to understand relations spatially. (But we needs to think of how to use them to their full potential)
  • Potential to scale structural attributes of forests is limited

  • But some future technological improvements are expected (GEDI, biomass satellite)

Thanks!

http://loicdutrieux.com/talks/ROBIN-2015-scaling