Environmental controls and predictions of African vegetation dynamics Martin Jung, Eric Thomas...
-
Upload
godwin-mitchell -
Category
Documents
-
view
214 -
download
0
Transcript of Environmental controls and predictions of African vegetation dynamics Martin Jung, Eric Thomas...
Environmental controls and predictions of African vegetation
dynamics
Martin Jung, Eric Thomas
Department of Biogeochemical Integration
Africa
• 2nd largest continent (30 x 106 km2)
• Lots of people (~1 billion)
• Comparatively little known
• All about water
• hyper-arid to tropical climate
• Hot-spots of interannual variability
• Vulnerable to climate change
Research questions
• Can we predict (forecast) seasonal and interannual vegetation dynamics?
• Which factors control vegetation dynamics (and where)?
• Can we generate an objective functional classification of the African vegetation?
• What causes large interannual variability?
Approach
Meteorology(7 x 4 + 7 x
24 )
Land use (8)
Soil (10)
Remotely sensed fAPAR
Remotely sensed fAPAR
Mean annual
Mean seasonal
cycle
Anomalies
Raw
Random forests
Lag
Cumulative Lag
Lag
Cumulative Lag
Variable selection based on Genetic Algorithm
Data & Methods
• Vegetation state = f(climate, land cover, soil)• Vegetation state: monthly FAPAR (1999-2009) from
SeaWiFS/MERIS (Gobron et al 2006, 2008)
• f: Random Forrests algorithm (Breimann 2000)
• Variable selection: Guided hybrid genetic algorithm (Jung & Zscheischler 2013)
• Climate: ERA-Interim (bias corrected), TRMM (rainfall)• Land cover: SYNMAP (Jung et al 2006) + FAO based land use
(Ramankutty & Foley 1999, updated)
• Soil: global harmonized world soil data base• Fire: GFED (Van der Werf et al)
Variables
• Climate: Tmin, Tmax, Precip, WAI, Rh, Rg, PET– Normal, mean annual, mean seasonal cycle,
anomalies– For normal and anomalies lag variables upto a lag of
6 months: lag, cumulative lag
Land use fractions: evergreen forest, deciduous forest, shrub, C3 grass, C4 grass, C3 crop, C4 crop, barren
• Soil: sand, silt clay, plant awailable water, Corg• Elevation, burned area
Experimental set-up
Variable selection using GHGA based on 500 randomly chosen locations
Training period: 1999-2004; Validation period: 2005-2009;
Leave ‘one year out’ forward run using selected variables (1999-2009);
20 Random Forests with 48 trees each using 1000 random locations
Evaluation of predicted fAPAREstimation of variable importances
ResultsOverall MEF = 0.91
Approach fails in some locations of massive transformations
MEF low, RMS high
MEF high, RMS low
MEF low, RMS low
MEF intermediate, RMS intermediate
Color composite of MEF and RMS
A little excursion…
Simple model based on soil moisture indicator explains 79%
of variance
Very small effect of fire on FAPAR anomalies
Back to the original model…
Local variable importance (sensitivity)
A functional classification
RGB of first 3 PCAs of variable importance (77% of variance
explained)
K-means clustering of variable importance (10 classes)
Just climate discriminates the groups!
* Groups = f(land cover, soil, climate)* 59 candidate predictors* Stratified random sampling (100 per class)* 6 variables selected (Overall accuracy of 78%)
Nor
mal
ized
var
iabl
e im
port
ance
What controls spatial pattern of interannual variability?
* STD(FAPARAnomalies) = f(land cover, soil, climate)* 59 candidate predictors* Training on full domain* 9 variables selected (MEF=0.82)
… again just climate!N
orm
aliz
ed v
aria
ble
impo
rtan
ce
5-1525-3545-5565-7585-95
Percentiles std(FAPARANO)
FPAR IAV high when:Intermediate WAI seasonality + Always high air humidity + Large IAV in radiation(but only part of the story!)
Outlook
• Potential of seasonal forecasting of FAPAR for early warning systems
• Long-term historical and future changes in FAPAR dynamics (e.g. changing patterns of distribution of functional groups, IAV)