Modelling the seasonal dynamics of Aedes albopictus populations using a spatio-temporal stacked machine learning model

Abstract

Various modelling techniques are available to understand the temporal and spatial variations of the phenology of species. Scientists often rely on correlative models, which establish a statistical relationship between a response variable (such as species abundance or presence-absence) and a set of predominantly abiotic covariates. The choice of the modeling approach, i.e., the algorithm, is itself a significant source of variability, as different algorithms applied to the same dataset can yield disparate outcomes. This inter-model variability has led to the adoption of ensemble modelling techniques, among which stacked generalisation, which has recently demonstrated its capacity to produce robust results. Stacked ensemble modelling incorporates predictions from multiple base learners or models as inputs for a meta-learner. The meta-learner, in turn, assimilates these predictions and generates a final prediction by combining the information from all the base learners. In our study, we utilized a recently published dataset documenting egg abundance observations of Aedes albopictus collected using ovitraps. and a set of environmental predictors to forecast the weekly median number of mosquito eggs using a stacked machine learning model. This approach enabled us to (i) unearth the seasonal egg-laying dynamics of Ae. albopictus for 12 years; (ii) generate spatio-temporal explicit forecasts of mosquito egg abundance in regions not covered by conventional monitoring initiatives. Our work establishes a robust methodological foundation for forecasting the spatio-temporal abundance of Ae. albopictus, offering a flexible framework that can be tailored to meet specific public health needs related to this species.

Keywords: Arthropod; Forecast; Invasive species; Mosquito; Population dynamics; Time-series..

Fig. 1

Biogeographical regions of Europe according to Cervellini et al. and the location (green dots) of the egg observations available in VectAbundance v0.15. The black lines represent the borders of the administrative areas of the countries of interest at the NUTS2 level. The map was created using R v4.3.

Fig. 2

(a) A conceptual representation of the stacking approach; (b) Framework of the modelling approach presented in the study.

Fig. 3

Median and interquartile range of the number of eggs observed (grey lines) and predicted by the regression model in both the internal and external validation. Both the observed and predicted values were aggregated over the three biogeographical regions to allow an easier representation.

Fig. 4

Median number of eggs predicted weekly by the regression model in the area of interest for the year 2022. The black lines represent the borders of the administrative areas of the countries of interest at the NUTS2 level. The grey areas are outside the area of interest. The map was created using R v4.3.

Fig. 5

Spatial representation of the predicted critical period-over-threshold (POT) for the year 2022 in the area of interest. White pixels are characterised by an average median weekly number of eggs always lower than 55. The black lines represent the borders of the administrative areas of the countries of interest at the NUTS2 level. The map was created using R v4.3.

Fig. 6

Modelled relationships between the Period-over-threshold and the interaction between the year and biogeographical regions using a Poisson GLMM.