Modeling Vegetation Cover at Sub-Pixel Scales for the Next Generation of Earth Remote Sensing Satellites

Summary:

• Landsat series of satellites: continuous set of space observations 1972 - Present

  • Wavelengths covered by different satellites overlap over time

  • The coverage has stayed relatively stable, not expanding much over the decades

• Differentiating green vegetation (GV) vs non-veg (NV)

  • Vegetation is much more reflective > 700 nm, much less < than non-veg

  • Can model any pixel as a mixture of veg / non-veg

  • Challenge: how to identify vegetation that is not clearly green (e.g. dead or senesced)

    • “Non-photosynthetic vegetation” (NPV)

    • Examples: Stems, branches, Dry grass, Crop residue, vegetation infected by pests

    • Varies seasonally

    • Indicator of wildfire danger

  • The spectra from the Landsat satellites are not great at differentiating non-veg from NPV

  • Need near infra-red

  • Goal: develop ways to identify NPV using hyperspectral imagery and ML

    • Upcoming satellites will have broader spectral sensors

• Upcoming Satellites (~2030 Launch):

  • SBG VSWIR: https://database.eohandbook.com/database/missionsummary.aspx?missionID=974

  • Landsat Next: https://landsat.gsfc.nasa.gov/satellites/landsat-next

• Data:

  • Collect on-ground spectral data of spectral absorption from different plant types

    • 12NPV, 6GV, 228 soils

  • Created spectrum mixtures to indicate what they would look like to a satellite with a given set of spectral sensors

  • Research question: What spectral info do we need to model fractional cover of NPV vs GV and soils?

    • 2030nm would be optimal

    • However, this wavelength is absorbed by atmospheric gases (CO2), which creates noise in the measurements

    • 2038nm - 2050nm is an open window that allows NPV to be detected 

    • Landsat Next mission is expected to have this sensor

      • Also provides continuity with Landsat 8/9 and 7, which have spectra that overlap Landsat Next but coarser

• How accurately can we model GV/NPV/soil using ML?

  • Focus on bands from Landsat Next sand SBG VSWIR

  • Trained many ML models, with Random Forest trained on hyperspectral synthetic mixtures achieving 15% RMSE on fractions cover (10% is the goal for highly usable prediction)

  • Analysis of model indicates that predictions for different cover types: GV/NPV/Soil includes features relevant to other types

  • Model is very accurate for GV; NPV and Soils are more challenging

  • Univariate models have similar accuracy to multi-variate models

  • Model applied to ground data

    • Much more variable: e.g. more samples that are purely Soil, GV or NPV

    • Much more challenging but error is still <10%

    • Observation: multi-variate modeling is much more successful at ensuring that the GV, NPV and Soil add up to 1

• Need much better field data

  • Drones, etc. (at 1.5cm resolution)

  • Goal: classify crop types

  • Leveraging hyperspectral aerial observations for mineral exploration