1. Land Data Assimilation
Tristan Quaife, Emily Lines, Philip Lewis, Jon Styles.

2. Last 6 month highlights
Implemented vertical heterogeneity in vegetation structure for land surface model RT schemes and observation operators
Implemented a particle filter for JULES

3. Canopy structure Task 2.2: Vegetation Structure
Task 2.3: Optical RT modelling
Canopy structure

4. Typical observation operator
1D-RT model of the canopy
Very simple canopy structure: Vertical homogeneity in leaf size, arrangement and reflective properties

5. PROSAIL Combines 4-stream canopy model SAIL
Calculates the diffuse and direct reflectance and transmittance of the whole canopy using:
Solar/viewing angle
Leaf area index (m2/m2)
Leaf angle distribution
Soil reflectance
Leaf reflectance/transmittance
(Verhoef et al. 2007)
with leaf optics model PROSPECT
Calculates the reflectance and transmittance of a single leaf using a plate model dependent on:
Internal leaf mesophyll structure
Chlorphyll a+b and carotenoid content (μg/cm2)
Dry matter content (g/cm2)
Equivalent water thickness (cm)
Brown pigment
– canopy RT model: no vertical variation in structure
(Jacquemoud & Ustin 2008)

6. Factors affecting reflectance
Leaf chlorophyll concentration
Photosynthetically active radiation (PAR) nm
Leaf area index (LAI)
Leaf angle
Simulations using PROSAIL

7. Observed vertical structure
Assuming vertical homogeneity is often not valid for real canopies:
Within-crown measurements from a temperate evergreen broadleaf species
Coomes et al. 2012
Leaves are often more upright at the top of the canopy and flatter at the bottom
Higher proportion of LAI found higher in the canopy, and leaves have higher mass/unit leaf area (LMA)
Whole-stand measurements from a temperate evergreen broadleaf forest
Holdaway et al. 2008
Whole-stand measurements from an temperate broadleaf forest
Wang & Li 2013
Leaf chlorophyll and water concentrations highest at the top of the canopy

8. Multi-layered PROSAIL
Canopy structural properties and leaf optical properties are constant within a layer
Properties vary between layers to represent vertical heterogeneity
To allow simulation of different reflective properties at different heights within a canopy, create a model consisting of layered slabs of homogeneous canopy using SAIL structure.
SOIL

9. Multi-layered PROSAIL
z=0
z=-1
Tu,1
Td,1
Td,2
Rt,2
Rb,1
Rt,1
layer 1
layer 2
Reflectance/transmittance of two layers combined:
𝑇 𝑑 ∗ = 𝑇 𝑑,2 𝐼− 𝑅 𝑏,1 𝑅 𝑡,2 −1 𝑇 𝑑,1
𝑅 𝑏 ∗ = 𝑅 𝑏,2 + 𝑇 𝑑,2 𝐼− 𝑅 𝑏,1 𝑅 𝑡,2 −1 𝑅 𝑏,1 𝑇 𝑢,1
𝑅 𝑡 ∗ = 𝑅 𝑡,1 + 𝑇 𝑢,1 𝐼− 𝑅 𝑡,2 𝑅 𝑏,1 −1 𝑅 𝑡,2 𝑇 𝑑,1
𝑇 𝑢 ∗ = 𝑇 𝑢,1 𝐼− 𝑅 𝑡,2 𝑅 𝑏,1 −1 𝑇 𝑢,2
Combining reflectance and transmittance of multiple layers

10. Vertical variation in leaf angle
homogeneous canopy structure
decline in leaf angle with height
Diffuse fluxes only. Simulation results show effect on canopy reflectance* of vertical variation in leaf area and angle compared to homogeneous canopy with same total LAI (3), same average leaf angle (40°)
Top of canopy
Bottom of canopy

11. Variation in leaf chlorophyll
homogeneous canopy structure
decline in leaf chlorophyll with height
Simulated effect on canopy reflectance of vertical variation in leaf chlorophyll concentration compared to homogeneous canopy with same average concentration (80μg/cm2)
Top of canopy
Bottom of canopy
Small decrease in reflectance in PAR region

12. Does this matter for LS models?
fAPAR is key biophysical variable for calculating primary productivity
Vertical structural heterogeneity affects light levels through the canopy
Land surface schemes (e.g. JULES) typically account for variable nitrogen, but not leaf angle or pigment properties

13. Da assimilation with jules
Task 2.1: Process model development
Da assimilation with jules

14. JULES

15. JULES: Carbon Budget

16. Fluxnet

17. Flux tower observations

18. Resampling Particle Filter
We have implemented a resampling particle filter for JULES
Uses the Metropolis-Hasting’s algorithm to perform the resampling
Implementation is very flexible
Requires no modification to the JULES code
Easy to adapt for different observations and different model configurations

19. Stochastic forcing Add noise into desired state vector elements
In following examples:
Daily stochastic forcing (JULES time step = 30min)
Truncated normal distribution
Soil carbon
Soil moisture (4 vertical levels)
Easy to change all of the above characteristics

20. Resampling step α = min 1, Loop over all particles, x
x* = random particle
y = observations
L(y|x*)
α = min 1,
L(y|x)
Draw z from U(0,1)
x* if z≤α
x if z> α
x =

21. Particle Filter

22. Non-assimilated variables

23. Pros/Cons Pros: Fully non linear Robust to changes in JULES
Easy to switch to other analysis schemes
e.g. Ensemble Kalman Filter
Cons:
Slow: approx 5 mins/particle/year
but algorithm is inherently parallelisable

24. Next 6 months

25. Immediate
Finish experiments on vertical structure and implement in JULES
Write up JULES Particle Filter experiments with Fluxnet data
Initial experiments against EO data

26. Next 6 months
Further modify JULES Sellers scheme to predict viewed crown and ground (for assimilation of long wavelength data)
Build 2-stage Data Assimilation algorithm:
EOLDAS for Leaf Area temporal trajectory and other slow processes (optical data)
Particle Filter for assimilating observations related to diurnal cycle (thermal, passive microwave)

27. EOLDAS & JULES phenology
JULES phenology routine is effectively separate from the rest of the model
Used to prescribe LAI profile, but not influenced by other parts of the model state
Consequently can be optimised stand-alone
Ideal application for EOLDAS
Use modified Sellers scheme as observation operator