3. Content
Satellite and meteorological data
• Satellite data sources
• Meteorological data sources
• Static data sources
Pre-processing
• Cloud masking
• Surface Albedo
• Thermal sharpening
• Smoothing
Biophysical Models
• Soil moisture algorithm - Trapezoid
• Evaporation, Transpiration and Interception (ETLook)
• NPP model – C-fix
Production and Delivery
4. Main Problem v2
Mismatch resolution NDVI and
LST is causing wrong values for
root zone soil moisture
5. Satellite data
• VIIRS Thermal infrared data (375 m) v3
– Brightness temperature
• Sentinel-2 optical and infrared data (10m / 20m)
– NDVI
– Albedo v3
– Indices v3
• Landsat 8-9 optical and thermal infrared data (30m /100m) NDVI
– Albedo v3
– Indices v3
• MSG SEVIRI data (3km)
– Atmospheric Transmissivity
6. Cloudmasking
• Sentinel-2 cloudmasking procedure
– Kappamask https://github.com/kappazeta/km_predict
– Sen2Cor mask
Kappamask is in most cases superior to Sen2Cor mask, however for
specific cases Sen2Cor is providing better results.
8. Cloud masking of inputs
Standard Sentinel-2 mask (sen2cor)
is a bit conservative (does not filter
all clouds and shadows) which
degrades the outputs
KappaMask is more rigorous in
masking, which results loss of data
(buffering), takes more time and
resources, but smoothened outputs
are better
instantaneous NDVI image of October2022, Mozambique (36KXC)
white spots = Sen2cor
grey scale = KappaMask
colours = cloudfree NDVI
Sen2cor masking KappaMask
9. Cloud masking of inputs
sen2cor KappaMask
SmoothenedNDVI image of January 2022, Yemen (38PMV)
10. Static data
• Copernicus elevation data v3
– Elevation
– Slope
– Aspect
• Copernicus vs WorldCover v3
– Maximum obstacle height
– Bulk stomatal resistance
• Statics update v3
– Longwave radiation parameters
14. Meteorological data
• (ag)ERA5 meteorological data v3
– Air and Dewpoint Temperature
– Wind speed
– Daily (agERA5) and instantaneous (ERA5)
– Used for final processing
• GEOS-5 meteorological data
– Daily and instantaneous
– NRT availability
– Specific humidity instead of dewpoint temperature
– Aerosol optical depth
15. ERA5/AgERA5
Daily AgERA5 and hourly ERA5 meteorologicaldata replace GEOS-5 for final
processing
ERA5 and AgERA5 provide more realistic results comparedto GEOS-5 but
differencesare relatively small:
• ERA5 temperature higher in humid tropics and lower in arid zones. Differences highest
in southern Africa and generally <3 Kelvin
• ERA5 windspeed lower in Sahel belt, higher in humid tropics. Differences highest in
south Africa and generally 1 m/s
• The relative humidity patterns differ significantly by season
16. Comparison to meteo stations (air temperature)
WASCAL
TAHMO ERA5
ERA5
GEOS-5
GEOS-5
ERA5 vs GEOS-5
ERA5 vs GEOS-5
R=0.71 R=0.60 R=0.76
R=0.85 R=0.84 R=0.85
18. Pre-processing
• Atmospheric correction v3
– VIIRS brightness temperature to land surface temperature (LST)
• Thermal sharpening v3
– pyDMS
• Smoothing timeseries
– Whittaker smoothing v3
– NDVI, Albedo and soil moisture
19. From brightness temperature to LST
Top of atmosphere measurementsmust be corrected for atmospheric and emissivity
effects to convert brightness temperatures into LST.
Split-window methods apply to sensors with at least two spectral channels. Sensors
with a single channel in the TIR domain (VIIRS has only one channel at 375m) need a
simulation of the atmospheric effects from an estimation of the atmospheric water
vapour and air temperature profiles. In such cases, Single Channel (SC) techniques are
required.
JPL has just developed a NRT LST algorithm for the 375 m VIIRS I5 band. Product is
available for a limited amount of time (approximate 10 days)
No funding for historical archive of 375m VIIRS LST product
20. Single Channel Algorithm (Munez et al., 2009)
Same firstprinciples:
• Planck's Law.
• Also here the atmospheric parameters of
incoming,outgoing radiation and
transmissivity need to be known
21. Single Channel Algorithm (Munez et al., 2009)
• The authors developed a solution where the atmospheric
water vapor content can be used to estimate the atmospheric
parameters using a second degree polynomial fit based on
simulations with a radiative transfer model (MODTRAN)
w = atmospheric water vapor
which we can get from GEOS-5 and is already used in the
computation of the clear-sky radiation in the soil moisture
algorithm
22. VIIRS LST
Options:
1. Use the 750m LST product for final and the VIIRS 375m LST NRT product (available on the
NASA LANCE near-real-time (NRT) system: https://nrt4.modaps.eosdis.nasa.gov/archive/allData/5200/VNP21IMG_NRT)
2. Replicate the VIIRS 375 LST NRT product, using the current emissivity calculation in
combination with a single channel atmospheric correction.
24. Thermal sharpening
• Thermal sharpening v3
– pyDMS
– Features used:
• Indices based on Sentinel-2 data
• Sentinel-2 bands
• Elevation features
• Regression tree with linear regression for each leaf
25. Soil moisture content – 6 October 2019
Underestimation of fields
due to different
resolution NDVI and LST
Large scale trend is fine, details may be wrongly interpreted
26. Irrigated fields – LST v2 vs v3
v2 (based on bilinear
resampled LST)
v3 (based on thermal
sharpened LST)
27. Soil moisture content – 6 October 2019
More logical soil
moisture values of fields
Large scale trend is similar to previous image, details are
better represented.
30. • Sentinel-2 Bands 2 and Band 8 (Blue and NIR)
• Elevation related features
– Slope
– Aspect
– Elevation
• Sentinel-2 based indices:
– MNDWI (Modified Normalized Difference Water Index) (SWIR1, green)
– NMDI (Normalized Multiband Drought Index) (NIR, SWIR1, SWIR2)
– VARI_RED_EDGE (Visible Atmospherically ResistantIndex Red Edge) (blue, red edge,
red)
– BI (bare index) (NIR, SWIR2, Red, Blue)
– PSRI (plant senescence reflectance index) (blue, red, red edge)
• In total more than 50 features have been considered for use
Features
32. Pixel sampling
Based on Coefficient of Variance of
multiple input features.
This is the CV of B8 of Sentinel-2
Pixels with low CV are considered
homogeneous and will be sampled
for the regression
35. Global vs Local Regression
Regression takes place in moving windows (local
regression). For the example this 5*5 = 25 windows
And for the whole image (global regression)
Local regression Global regression
39. Result (without residual correction)
This image is correctedagain for
residuals based on a comparison
with the resampled low resolution
LST, to correct biases in the result
51. Root zone soil moisture
• Trapezoid
• Pixel-by-pixel solution (not based on image statistics)
• Penman-Monteithfor extreme dry edge
• Wet bulb temperature and air temperature for wet edge
• Free convection at low wind speeds v3
52. Soil moisture
• Points A and B are
calculated (for each
pixel!) based on
Penman-Monteith
• Points D and C are
provide by the air
temperature
• For WAPOR-ETLook
version 2: Point D is
provided by the wet
bulb temperature
53. Soil moisture parameterization (aerodynamic resistance)
Windspeed of GEOS-5 and ERA5 gets unrealistically low in certain conditions => if the
surface heat flux becomes sufficiently high, it will generate turbulence (e.g. wind)
which provides a negative feedback on that heat flux and surface temperature. This
feedback was missing, leading to conditions with a very low windspeed (<1 m/s) in
combination with a very high heat flux, resulting in unrealistically high (dry) surface
temperatures
Solution is to calculate the
aerodynamic resistance for free convection
independent of wind or friction velocity:
The aerodynamic resistance feeds into the soil moisture model
Impact is relatively small (only happening is specific areas)
54. Soil moisture parameterization (aerodynamic resistance)
Ra forced convection
Ra free convection
Minimal value of
Ra is chosen
Impact is that
maximum LST gets
lower and that is will
become drier sooner
(e.g. in desert areas)
Procedure to calculate the bare soil maximum temperature
(part of trapezoid)
58. ETLook
• Penman-monteith solution for evaporation and transpiration
• Interception is modelled separately, energy is subtracted
• Surface resistance is modelled using Jarvis approach with four
separate stress factors:
– Air Temperature
– Vapor Pressure Deficit
– Solar Radiation
– Root zone soil moisture
• Soil resistance is related to soil moisture
59. Theory - ETLook
• Penman monteith equation
• Solved separately for two components
– Evaporation (soil)
– Transpiration (canopy)
• Interception
• Daily timestep
• Two soil moisture layers (topsoil / subsoil)
• ETLook paper WRR Bastiaanssen et al. (2012)
60. Penman-Monteith
Canopy Transpiration
• 𝑄𝑐𝑎𝑛𝑜𝑝𝑦
∗
= 1 − 𝛼0 𝑆↓
− 𝐿∗
− 𝐼 1 − exp −0.6𝐼𝑙𝑎𝑖 Net Radiation - Canopy
Soil Evaporation
• 𝑄𝑠𝑜𝑖𝑙
∗
= 1 − 𝛼0 𝑆↓
− 𝐿∗
− 𝐼 exp −0.6𝐼𝑙𝑎𝑖 Net radiation - Soil
Total net radiation
Total net radiation
Extinction function
Extinction function
61. C-Fix
• C-fix computes the Net Primary Production
• C-Fix is a Monteith type parametric model driven by
temperature, radiationand fraction of Absorbed
Photosynthetically Active Radiation (fAPAR)
• fAPAR is determined by NDVI
• Soil moisture stress is taken into account (similar to ET)
62. Production
• Use of MGRS tiling system v3
• AWS cloud computing v3
• OpenDataCube for registering geodata v3
64. AWS cloud computing
• Change from on-premise computing (private cloud) to AWS
• Change from Airflow to Dagster for scheduling processes
• Subdivisioninto tiles makes parallel processing possible