This repository contains the workflow pipeline for training and inferencing the Climademic Suitability Model, as described in the article. The model combines One-Class Support Vector Machines (OC-SVM) with an incremental learning strategy to continuously update suitability predictions using newly available mosquito occurrence observations.
The workflow begins with the preparation of all datasets required for training, incremental learning, and global inference:
- Initial training dataset – generated from the Global Compendium occurrence records by extracting climate, land-use, and population variables for the corresponding geographical masks.
- Global inference dataset – generated by dividing the global land surface into 0.25° × 0.25° grid cells and extracting environmental predictors for each cell. Climate variables are extracted for all 12 months of the year, allowing month-specific suitability predictions. The dataset is generated on a yearly basis, with a separate file produced for each year.
- GMOD dataset – generated from GMOD mosquito observations by extracting the corresponding environmental predictors.
Important: The global inference dataset must be prepared before generating the GMOD dataset. In the current implementation, land-use and population variables for GMOD observations are obtained from the global inference dataset. Specifically, each GMOD observation is assigned the land-use and population values of the 0.25° × 0.25° grid cell in which the observation is located. Therefore, the global dataset serves as the source of these variables during GMOD dataset creation.
Once all datasets have been prepared, the initial OC-SVM model can be trained using the Global Compendium dataset and subsequently updated through the incremental learning workflow using GMOD observations.
The initial OC-SVM model is trained using historical environmental data consisting of:
- Climatic variables
- Land-use data
- Population density data
These variables are extracted using geographical masks derived from Global Compendium occurrence polygons spanning 1975–2014. Polygon sizes range from 5 × 5 km to 110 × 110 km. For climatic variables, the 75th percentile (0.75 quantile) is used to characterize environmental conditions within each polygon.
The The resulting model represents the climatic and environmental niche associated with known mosquito occurrences up to the year 2014.
After the initial training, the model is updated annually using contemporary mosquito observations from the GMOD dataset.
The update procedure is as follows:
- Train the baseline OC-SVM model using Global Compendium data up to year t (e.g., 2014).
- Extract environmental features for GMOD observation locations for year t + 1 using a spatial resolution of 0.25° × 0.25°.
- Since GMOD observations are reported on a daily basis, the exact month of observation is known. Climatic predictors corresponding to the observation month are therefore used when generating model inputs.
- Apply the current model to predict suitability for all GMOD-derived locations.
- Locations with observed mosquito presence that are predicted as unsuitable (suitability score < 0.5) are treated as misclassified observations.
- These misclassified observations are added to the training dataset.
- The OC-SVM model is retrained using the expanded dataset.
Steps 2–7 are repeated for each subsequent year until all available GMOD data have been incorporated.
Forr model inference, the global land surface was divided into 0.25° × 0.25° grid cells, and a feature dataset was generated for each cell using the corresponding environmental predictors.
The model produces two output scores for each grid cell:
- Prob_1 – the suitability score, representing the likelihood that the environmental conditions are suitable for mosquito occurrence.
- Prob_2 – the probability that the area is unsuitable for mosquito occurrence.
These outputs can be used to generate global suitability maps and to classify areas as suitable or unsuitable based on a predefined threshold (e.g., 0.5).
The model should be trained separately for each mosquito species (Aedes aegypti, Aedes albopictus)
Each species-specific model is trained and updated using occurrence records and environmental data corresponding only to that species, allowing the model to capture differences in their ecological niches and environmental suitability requirements.
The current implementation is a collection of procedural scripts and has not yet been refactored into classes or packaged as an installable Python package.
- GMOD file depending on a global dataset
- No explicit choice of the grid limits (now its only [-180; -90: 180, 90])