Spatial Statistics 2017 Conference: Towards the modelling and mapping of active questing ticks using random forest
1 like114 views
The document discusses a study on modeling and mapping questing ticks using random forest techniques, in response to the rise of vector-borne diseases due to various global changes. It details the methodology, which includes multivariate regression and a data-driven approach to predict tick dynamics and assess feature importance, highlighting key contributors like evapotranspiration and humidity. The research also emphasizes the potential of citizen science in gathering environmental data for scientific analysis.
Spatial Statistics 2017 Conference: Towards the modelling and mapping of active questing ticks using random forest
- 1. Towards the modelling and mapping of active questing ticks using random forest Irene GARCIA-MARTI Raúl ZURITA-MILLA Session B3: GeoHealth Spatial Statistics Conference 5th July 2017
- 2. Context Re-emergence of vector-borne diseases: Global & socio-economic changes VBD: 23 diseases (including Zika) Changes + ticks: Longer tick season Increase in tick populations Northward + elevation expansion New suitable habitats Source: WHO
- 3. Data Since 2006: Trained volunteers sample 17 locations in NL on a monthly basis Count ticks in its different life stages (i.e. larvae, nymph, adults) First citizen science project of its kind! Source: WUR 3
- 4. Data Habitat: • Coniferous • Deciduous • Bushlands Questing ticks in the day: • 2006 – 2014 • ~3,000 observations • Two transects per site
- 5. Data
- 6. Data
- 7. Data 7 From “Lyme disease: The ecology of a complex system” R.Ostfeld (2012)
- 8. Objectives Multivariate regression with Random Forest: Goals: Predict tick dynamics in space and time Assess feature importance to predict tick dynamics Data-driven approach: Does not assume any distribution Fits non-linear natural processes Deals with high-dimensionality
- 9. Methodology • T, P, EV, RH, SD, VP • KNMI, 1km, daily Weather • Veg. indices • MODIS, 500m, 8- day Remote sensing • Tick habitat, mast years, land coverEcology
- 10. Methodology 2. Modelling w/ Random Forest: Ensemble learning model Forest of DT Trained w/ bootstrap samples Characteristics: Perturbation: Ensure diversity Trees have high variance Random selection of features Variable importance Tree T Forest of decision trees
- 11. Methodology • Weak: DT • Strong: Bayesian Type of learner • Bagging • Boosting Growing scheme • Algebraic combiner • Voting-based method Combination rule + RSF = Random Forest Ensemble Learning
- 12. Methodology 1) Draw a random bootstrap sample from the dataset 2) Grow a tree estimator on the selected samples Select n features at random Find the best split on the selected n features Grow the tree estimator: child nodes 3) Repeat 1-2 for a number of tree estimators 4) Combine the predictions by averaging bagging DT RSM splitting branching combining
- 13. Methodology In general, ML algorithms do not perform well with time: Multiple weather conditions yield same counts of tick activity Hampers learning process of the algorithm 3. Temporal Z-scores: Re-scaling the dependent variable Descriptive statistics in the long time-series to find the range Trees can distinguish between conditions that yield low tick activity
- 14. Results 14
- 15. Results Top ten most important features: Percent of tick activity captured by each feature alone Short-term vs long-term weather events Atmospheric water levels seem to be prominent to predict tick activity # Feature Time unit Importance 1 Evapotranspiration 365 15% 2 Relative humidity 30 11% 3 Min. Temp. 365 7% 4 Evapotranspiration 2 5% 5 Precipitation 3 4% 6 Evapotranspiration 5 3% 7 Relative humidity 365 3% 8 Min. Temp. 90 2% 9 Precipitation 365 2% 10 Land Cover - 2%
- 16. Results 16 Predicted tick activity June 1st, 2014 Trained model is applied to each pixel with forest of the Netherlands Interpretation: Provinces of Drenthe and Groningen presented high activity of ticks on that day Randstad area presented the lowest tick activity
- 17. Results 17
- 19. Conclusions Case of data fusion of volunteered data and environmental data: Volunteered dataset is very noisy, but we could model transects with high and low activity together Predict tick activity in space and time Assess the feature importance to predict tick activity Added time-awareness to a regression with Random Forest Citizen science initiatives allow the crowdsourced monitoring of environmental phenomena and can produce geospatial data collections to serve scientific workflows
- 20. Thanks for your attention