Data Assimila*on of Satellite Ac*ve Fire Detec*on in Coupled Atmosphere-‐
Fire Simula*on by WRF-‐SFIRE
Jan Mandel University of Colorado Denver
with Adam K. Kochanski, Martin Vejmelka, Jonathan D. Beezley,
and Sher Schranz University of Utah, Czech Acadeny of Sciences, Kitware,
Colorado State University/NOAA
Supported partially by NASA NNX13AH59G, NSF DMS-1216481, and Czech Science Foundation 13-34856S.
VII Interna)onal Conference on Forest Fire Research, Coimbra, Portugal, November 18, 2014
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!SFIRE
Atmosphere!model!WRF
Surface!fire!spread!model
Wind
Heat!and!vapor!fluxes
Fuel!moisture!model
Surface!air!temperature,!rela?ve!humidity,rain
Chemical!transport!model!WRFBChem
Fire!emissions!(smoke)
RAWS fuel moisture sta)ons VIIRS/MODIS fire detec)on
HRRR forecast
Data assimila)on
WRF-‐SFIRE components
Data assimila*on
Satellite Fire Detec*on – 2010 Fourmile Canyon Fire, Boulder, CO
3
MODIS/VIIRS Ac*ve Fire Detec*on Data
• MODIS instrument resolu)on 750m at nadir to 1.6km, geo-‐loca)on uncertainty up to 1.5km, VIIRS resolu)on 375m.
• MODIS processed to 1km detec)on squares available, VIIRS 750m. Research processing VIIRS to 375m polygons exists (Schroeder 2013). Much coarser scale than fire behavior models (10-‐100m)
• False nega*ves are common. 90% detec)on at best. 100m2 flaming fire has 50% detec)on probability (MODIS. VIIRS is be[er but nothing can be ever 100% accurate).
• No detec)on under cloud cover -‐ should not count
• Detec)on squares in arbitrary loca)ons -‐ fire sensed somewhere in the square, not that the whole square would be burning.
• No global, binary fire/no fire map.
• Data assimila)on = data improve the model in a sta)s)cal sense.
5
• Ac)ve Fire detec)on should be used to improve fire modeling in a sta)s)cal sense only, not as a direct input.
• Data assimila)on ≠ cyclic reini)aliza)on from new data.
Assimila*on of Ac*ve Fire detec*on data
VIIRS Ac*ve Fire Detec*on for 2013 Barker Canyon fire
VIIRS fire detec)on squares
Simulated fire arrival )me Time
MODIS ac*ve fires detec*on with simulated fire arrival *me
7
A method for assimila*on of ac*ve fires detec*on data
• Modify the fire arrival 0me to simultaneously minimize the change and to maximize the likelihood of the observed fire detec;on.
• Inspired by computer vision in Microsoa Kinect, which modifies a model of human mo)on to simultaneously minimize the change and to maximize the likelihood of the observed images (A. Blake, Gibbs lecture at JMM 2014)
• Bayesian sta)s)cs view: maximum posterior likelihood, found by nonlinear least squares.
8
Fit the fire arrival )me T to the forecast Tf and fire detec)on data , inspired by computer vision (Microsoa Kinect design, Blake 2014)
9
Assimila*on of MODIS/VIIRS Ac*ve Fire detec*on
• Ts = satellite overpass )me • constraint C(T-Ts)=0 : no change of fire arrival )me at igni)on points • f(t,x,y) = likelihood of fire detec)on t hours aaer )me arrival • A =ellip)c differen)al operator to penalize non-‐smooth changes • smooth descent direc)on δ by solving the saddle point problem
J (T ) = ε
2T −T f
A
2− f (T S∫ −T ,x, y)dxdy → minC (T−T f )=0 ,
Aδ +Cλ = ∂
∂tf (T S −T ,x, y), CTδ = 0,
f(t,x,y) : log of the likelihood of fire detec)on as a func)on of the )me t elapsed since the
fire arrival at the loca)on (x,y)
10
Assimila*on of the VIIRS Fire Detec*on into the Fire Arrival Time for the 2012 Barker Fire
11
Forecast Search direc)on
Analysis Fireline = contour of fire arrival )me
Decrease of the fire arrival )me
VIIRS fire detec)ons
Time
Forecast fire arrival )me
But fire is coupled with the atmosphere
12
Atmosphere
Heat release
Fire propaga)on
Wind Heat flux
• Heat flux from the fire changes the state of the atmosphere over )me.
• Then the fire model state changes by data assimila)on. • The atmospheric state is no longer compa)ble with the fire. • How to change the state of the atmosphere model in
response data assimila)on into the fire model? • And not break the atmospheric model.
Spin up the atmospheric model aRer the fire model state is updated by data assimila*on
Fire arrival )me changed by data assimila)on
Ac*ve fire detec*on
Atmosphere out of sync with fire
Forecast fire simula)on
Coupled atmosphere-‐fire
Replay heat fluxes derived from the changed fire arrival *me
Rerun atmosphere model from an earlier *me
Con)nue coupled fire-‐atmosphere simula)on
Atmosphere and fire in sync again
Conclusion • A simple and efficient method – implemented by FFT 2-‐3 itera)ons are sufficient to minimize the cost func)on numerically, 1 itera)on already pre[y good
• Pixels under cloud cover do not contribute to the cost func)on
• Standard bayesian data assimila)on framework Forecast + data = analysis
• Future: – Ac)ve fire detec)on likelihood from the physics and the instrument proper)es?
– Combina)on with standard data assimila)on into the atmospheric model, e.g. add 4DVAR cost func)on?
14