Validation of an agent-based model of shifting agriculture
- 1. Validation of an Agent-based Model of Shifting Agriculture A village case study from uplands of Vietnam The An Ngo Linda See Frances Drake
- 2. Outline of the presentation Introduction to the Agent-based Shifting Cultivation Model (ASCM) General validation process Validation results (e.g. comparing model outputs and the real data) Conclusions
- 3. Introduction to the ASCM What is Shifting cultivation ? Why need a model? A swidden field after burning Crop at young stage at swidden fields Fallow… and forest degradation Slash and burn to clear the field Bare soil after harvest The VN government has implemented policies to stop shifting cultivation But Swidden is still wide spread Why ?
- 4. Structure of the ASCM The model consists of 3 components: Household agent Biophysical environment (land agent) Global parameters (policy, market price etc.) And several routines; the most important is: Land cover transition Global parameters (3) Markets, Government policies etc. Policy lever (Pressure) Land user (1) (Household agent) Land patches (2) (Land agent) Adaptation Tenure relations Sharing strategies Land use patterns
- 5. Land Cover Transition An framework of Updating vegetation transition Transition probability = f { Cover (t-1) , Clearing, Logging, Vegetation growth } Vegetation growth = f { fallow-age, soil, neighbour-patches } Dense forest with dominant species Clearing and Continuous Intervention Burning Selected logging Natural succession C5. Secondary/regenerated forest (Dense forest) C4. Bushes and scattered wood trees (Open forest) C3. Tall grasses and shrubs C2. Short grasses or bare land C1. Crops
- 6. The ASCM is validated at Binh Son-1 village
- 7. Validation Process Initial model 2. Calibration Selecting the range of values for the model parameters 1. Sensitivity Identifying significant parameters 3. Output validation Do predicted results match reality? Fully validated model
- 8. Data for Validation (Note: The starting year of the simulation is 2000) Validation process Data Measurement description (1) Sensitivity analysis Land cover map Simulated 2007 vs. satellite 2000 (2) Calibration (GAs) Land use, Land cover map Simulated 2005 vs. Satellite 2005 (3) Output validation: - Land cover change Land cover map Simulated 2006 vs. Satellite 2006
- 9. Sensitivity analysis using Sequential Bifurcation technique Validation Results Identifying parameter’s values lead to low and high land cover change Identifying the individual effect by switching on and off the parameters Factor or Parameter Low (0) High (1) Rice price (1000 VND/kg husk rice) 5 1 Yield threshold for field selection (ton/ha) 1000 600 Access to the agricultural training (%) 2 8 Access to the credit (%) 9 1 Subsidy per capita (1000 VND) 100 10 Building rate (%) 0.01 0.15 Equipment buying rate (%) 0.1 0.3 Allocated land per capita (ha) 1.406 0.469 Land per labour workability (ha) 0.3 0.5 Land per dependent workability (ha) 0.05 0.01 Daily food requirement (kg husk rice/capita) 500 7500 Forest protection pressure (scale 1-10) 9 1 Population growth rate (%) 2.5 1.3 Household growth rate (%) 2.5 1.3 Individual bebefit ( w 1 ) (scale 0.0 – 1.0) 0.1 0.9 Group bebefit ( w 2 ) (scale 0.0 – 1.0) 0.9 0.1 Individual awareness ( w 3 ) (scale 0.0 – 1.0) 0.9 0.1 Institutional influence ( w 4 ) (scale 0.0 – 1.0) 0.9 0.1 Description Effect Shrub transition threshold ( Shrub ) 16.71 Response to the agricultural policies ( r a ) 16.33 protection pressure 8.40 Land per labour workability 7.39 The Communist Party cell’ influence 7.11 Response to the forest policies ( r f ) 6.42 Average expected yield 5.86 Yield threshold for field selection 4.89 Equipment buying rate 3.51 Subsidy per capita 3.37 Daily food require 2.95 Group benefit ( w 2 ) 2.26 Tree transition threshold ( Tree ) 2.15 Population growth rate 2.12 Access to the credit 2.05 Household growth rate 1.99 Institutional influence ( w 4 ) 1.71 Village Chair Men’s influence 1.43 Access to the agricultural training 1.30 Individual benefit ( w 1 ) 1.13
- 10. Calibration using Genetic algorithm (GA) Optimal GA values selected for model operation Parameters (Range from 0.0 – 1.0) GA values Individual bebefit ( w 1 ) 0.3 Group benbefit ( w 2 ) 0.8 Individual awareness ( w 3 ) 0.8 Institutional influence ( w 4 ) 0.8 Response threshold of forest policies ( r f ) 0.6 Response threshold of agricultural policies ( r a ) 0.4 Tree transition threshold ( μ 0-tree ) 0.4 Grass-shrub transition threshold ( μ 0-grass&shrub ) 0.8
- 11. The arrangement of the evaluated window in the MRG test Fit estimated by MRG at multiple resolutions for the simulated and reference map of 2006 Output validation: Multi-resolution goodness-of-fit (MRG) validation of spatial patterns of land cover
- 12. Visual Comparison Between the Model Output and the Reference Data Reference map 2000 Simulated map 2006 Reference map 2006
- 13. Validation results indicate that: ASCM compares favourably with the data from the real shifting cultivation system at the study site Limitations of the model: Validation in the short-term (2000-2006) due to the availability of the real data Lack of multiple validations scenarios (e.g., validate for different villages or policy programmes). This will be a topic for another paper. Next steps: Apply the model to a series of scenarios to look at forest protection, population growth, etc. Conclusions/Next Steps
- 14. Thank you !
Editor's Notes
- #5: Several routines: three important ones will be described
- #6: Fuzzy logic for assessing transition probability (linear monotonically increasing membership) Identify the threshold value of the transition by GAs
- #8: This is the “structural validation” (Carley, 1996)
- #9: Data: depends on the availability of the data
- #10: Sequential Bifurcation (SB) estimates the sensitivity of the parameters in the model by: Identifying the values of the parameters that could result in low and high model outputs, Switching on (use the high values) and off (use the low values) to determine the different in the model outputs Estimating the factor effects using the linear regression model
- #11: GA: Calibrated parameters are randomly assigned values (within their ranges) and used to run the model. Model output are then compared with the real data. The value set of parameters that produce the highest accuracy will be selected by the GA
- #12: MRG: compare the structure land cover between 2 maps. This is undertaken at various resolutions; start by 1x1 pixel and increase the size until they cover all the map. This measurement takes into the consideration about the patterns of land cover (not the only number of pixel correct matched as the other methods do) MRG compares: - Null model: 2000 map (no change until 2006) with reference 2006 Complete random predicted map with reference 2006 Simulated map in 2006 with reference map (real data) higher than all others The acceptable Ft values from previous research often range from 7.0 – 9.0. The result (8.1) is between this range
- #13: Overall agreement: 2000 ref – 2006 ref = 45% 2006 sim – 2006 ref = 46% Final Ft: 81% --- 69%