Temporal modeling of surface water bacteriological quality in West Africa using remote sensing and machine learning methods

In 2021, diarrheal diseases were responsible for around 1.17 million deaths worldwide. Sub-Saharan Africa is one of the most impacted regions, where 440,000 deaths were recorded in 2024. This high mortality rate can be explained by 1) significant bacteriological pollution of surface waters by pathogenic micro-organisms responsible for diarrheal diseases, 2) widespread use of untreated water by the population and3) lack of sanitation and community health infrastructures. In addition, ongoing climate change is likely to have a negative impact on water quality, and potentially increase the presence and transmission of pathogens.

Tele-epidemiology, the combination of satellite observations and epidemiology, is a powerful tool for studying climate-environment-health relationships and for understanding and predicting the spatio-temporal distribution of pathogens and diseases through the use of satellite and in-situ data. We aim at using this method to indirectly monitor water quality and reveal environmental factors conducive to the emergence of critical health situations by modeling the dynamics of E. coli in West Africa. E. coli is considered the best indicator of faecal contamination (IFC) in temperate zones, and is recommended as a proxy for assessment of water contamination. In Burkina Faso, Robert et al (2021) demonstrated a significant correlation between E. coli, intestinal enterococci and cases of diarrhea. E. coli therefore appears to be a good IFC in West Africa, and would be relevant for predicting diarrheal diseases.

The first objective is to study the links between E. coli concentration in water and environmental parameters 1) measured in-situ in West African surface waters (Bagre reservoir in Burkina Faso and Kongou - Bangou Kirey in Niger) from 2018 to 2024 (concentration of suspended particulate matter, particulate organic carbon, etc.), 2) measurable by satellite (surface water reflectances mesured by Sentinel-2) or 3) estimable by satellite algorithm (precipitation, hydrometeorological parameters, NDVI, etc.). We then use key environmental parameters to model the concentration of E. coli in these sites over several years, firstly using all parameters, and then only using satellite data to study their robustness. Various machine learning models (Random Forest, SVM, KNN, etc.) were tested and compared with each other (calculation of R², RMSE, MSE and MAPE). 

For the Bagre site, the best model of E. coli concentration had showed a R² of 0.76 (RMSE 0.49 log10 MPN/100mL) using in-situ and satellite data, and R² of 0.69 with only satellite data (RMSE 0.56 log10 MPN/100mL). For Kongou, the best model had showed a R² of 0.7 using in-situ and satellite data, and R² of 0.65 with only satellite data.

This work will allow to create health hazard indices that can be used by public health players, firstly in West Africa without the need to collect data in the field, and then more generally for other sites facing similar public health problems.