Historical Time Series Biomass Modeling: To Train on Plots or Pixels?

AGU – Fall 2022


Lucas Johnson
Michael Mahoney
Colin Beier

Context & motivation

  • Forest carbon sink in NYS expected to double in the next 30 years.

  • Translating FIA information into maps.

  • Historical time series mapping spatially explicit stock change information.

Our approaches

Landsat figures adapted from eMapr content https://emapr.github.io/LT-GEE/landtrendr.html

One-staged

Two-staged

Our approaches

Landsat figures adapted from eMapr content https://emapr.github.io/LT-GEE/landtrendr.html

One-staged

Two-staged

LiDAR-AGB in NYS

  • Models developed with FIA and public leaf-off LiDAR.
  • 30 m AGB for ~60% of NYS.




What are the tradeoffs between one-staged and two-staged modeling?



Is there a clear winner?

Modeling framework

  • Two-staged – 20,000 pixels.

  • One-staged – ~2000 plots.

  • Landsat spectral indices, topo, climate predictors.

  • ML ensemble models:

    • Random forest
    • Gradient boosting machines
    • Support vector machines

  • Test-set accuracy vs. holdout panel accuracy.

Test-set accuracy

  • One-staged
    • 41% RMSE
  • Two-staged
    • 25% RMSE

Holdout panel accuracy

  • One-staged
    • 41% RMSE
  • Two-staged
    • 44% RMSE

Huntington Wildlife Forest

Huntington Wildlife Forest

Huntington Wildlife Forest

One-staged

  • Slightly more accurate against FIA plots.

  • More efficient.

Two-staged

  • Captures more spatial heterogeneity.
  • Less saturation.

Summary


  • Critical to validate two-staged modeling with field inventory.


  • Train on the distribution you want to predict.


  • How to make use of LiDAR information in this context?

Thank you


Access these slides at:

lucaskjohnson.com/agu-2022

Find me online at:

lucaskjohnson.com
twitter.com/lucaskjohnson03