Abstract
The global climate is undergoing relentless change. While it is a complex phenomenon with difficult-to-predict full-scale impacts, its influence on the emergence of diseases, especially vector-borne diseases (VBDs), is widely acknowledged. Therefore, comprehending our landscape and examining its potential shifts over time due to alterations in climatic and environmental factors becomes crucial.
In the framework of the WOAH project "Defining Ecoregions and Prototyping an EO-based Vector-borne Disease Surveillance System for North Africa (PROVNA)," we have identified ecoregions across North Africa, sharing similar environmental and climatic characteristics and which can serve as a pivotal foundation for enhancing surveillance systems and early warning system for vector-borne viruses.
Selected EO data products for the period 2018-2022 (Land Surface Temperature Day and Night - LSTD and LSTN, Normalised Difference Vegetation Index - NDVI, Soil Moisture - SM, Normalised Difference Water Index - NDWI, Rainfall - RF) at 250 meters/16 days resolution have been collected, aggregated, and standardized at a season/year level. We used the (super)SOMs method (an unsupervised neural network clustering method) to get a topology-preserving map, transforming a complex high-dimensional input space into a simpler low-dimensional (typically two-dimensional) discrete output space. The resulting map is easy to interpret and can be used for classifying new (both in space and time) observations.
A stratified random sampling, preserving the spatial and temporal variation of the environmental factors, was used to half the ~3 hundred million pixels of the entire extent and to create the dataset to train the SOMs.
An RGB colorization was applied to the trained (super)SOM map (a hexagonal grid made up of 40x40 neurons for 20,000 epochs) to further improve its readability with respect to the input environmental variables averages (the red channel was assigned to LSTD and LSTN, the green channel to NDVI and NDWI and the blue channel to RF and SM). Finally, the affinity propagation clustering algorithm was applied to the map’s neurons (1.600) to group them and get the number of distinct ecoregions.
The map clustering was in the end used to classify all of the pixels and to get yearly classification rasters.
The proposed approach allows to apply the SOMs algorithm to EO data for a very large area, classifying landscape (and its temporal evolution) while preserving high spatial resolution and at the same time getting an immediate interpretation of similarity. This model will effectively support Competent Authorities in North Africa to identify locations where applying surveillance activities (while optimising financial, material and human resources) for identifying potential viruses in the area such as Rift Valley Fever which pose a threat for the entire Mediterranean region, including Europe.