Tracking Malaria From Space: What Water and Altitude Reveal About Risk in Tanzania

Researchers studying Karagwe District in northwestern Tanzania explored whether satellite observations and machine learning could help explain differences in malaria prevalence across the landscape. Their findings put water-related conditions at the centre of the picture and expose important limits in predicting changes from one year to another.

Tracking Malaria From Space: What Water and Altitude Reveal About Risk in Tanzania
Representative Image Image Credit: Gemini

A stretch of wet ground or a low valley can reveal conditions that support malaria transmission, especially where regular field surveys are difficult to organise. Researchers studying Karagwe District in northwestern Tanzania explored whether satellite observations and machine learning could help explain differences in malaria prevalence across the landscape. Their findings put water-related conditions at the centre of the picture and expose important limits in predicting changes from one year to another.

The research, "Integrating satellite-derived environmental variables and machine learning techniques in predicting malaria prevalence in Karagwe District, Tanzania," was published in Frontiers in Artificial Intelligence. Joanitha Kaijage, Lawrence Mdegela, Devotha G. Nyambo and Elizabeth Mkoba examined environmental information and malaria estimates covering 2020–2024. Their work focused on the estimated prevalence of Plasmodium falciparum infection among children aged two to ten.

Karagwe's rural communities live among highlands, valleys, wetlands and lakes, creating varied conditions for mosquitoes and the people they bite. Reports of parasites showing partial resistance to artemisinin, a key malaria medicine, added urgency to the researchers' interest in the district. Drug resistance itself was outside the study's analysis, which examined environmental patterns associated with estimated malaria prevalence.

Reading the landscape through satellite data

The team combined information from Sentinel-2, MODIS, CHIRPS, the Shuttle Radar Topography Mission and the JRC Global Surface Water dataset. These sources supplied measures of vegetation, surface water, moisture, built-up land, rainfall, land surface temperature, elevation, slope and distance to water. Google Earth Engine helped process the observations into annual datasets aligned with a five-kilometre analysis grid.

Malaria information came from the Malaria Atlas Project, whose maps estimate parasite prevalence by combining survey observations with environmental and geographical information. The researchers were working with an existing model's estimates rather than collecting independent infection measurements from children in Karagwe. Their results describe how satellite observations relate to that estimated malaria surface, a distinction that matters when judging practical usefulness.

Several environmental measures carried overlapping information, so the researchers reduced the final set to water and moisture indices, temperature, rainfall, elevation and distance to water. Two machine learning approaches, Random Forest and XGBoost, searched for relationships that simple straight-line comparisons could miss. Testing included geographically separated areas and entire years withheld from training, allowing the team to examine how well each model handled unfamiliar data.

Water leaves the strongest signal

The Normalized Difference Water Index, or NDWI, emerged as the most influential predictor, accounting for 42.61% of the reported relative feature importance. Elevation and the moisture index each received 15.74%, followed by rainfall at 9.99%, land surface temperature at 8.76% and distance to water at 7.15%. These percentages describe importance within the model, rather than the share of malaria infections caused by each factor.

An explanation method called SHAP showed that higher NDWI values generally pushed predicted prevalence upward, with higher elevations tending to lower predictions. The water index enhances surface-water features, giving researchers a broad indication of local water conditions. Individual temporary mosquito breeding pools are too small to be reliably identified on the study's five-kilometre grid, so NDWI cannot be treated as a direct map of breeding sites.

Rainfall showed a negative relationship with estimated prevalence, a finding that needs more context than the idea that rain creates mosquito habitat. Heavy rain can flush out breeding sites, and annual summaries can hide seasonal changes or the delay between rainfall and malaria transmission. The authors discuss these possible explanations without establishing which mechanism produced the observed pattern in Karagwe.

Better at mapping places than predicting new years

XGBoost performed moderately under the main spatial evaluation, explaining roughly 55% of variation in the mapped prevalence estimates, compared with about 51% for Random Forest. Its root mean square error was 1.409 and mean absolute error was 0.994, slightly lower than Random Forest's corresponding results of 1.476 and 1.023. These scores came from testing across 54 geographical blocks, each measuring ten kilometres by ten kilometres.

A conventional random split produced much stronger scores, with XGBoost reaching an R² of 0.871 and Random Forest reaching 0.846. Nearby observations can share environmental characteristics, making a random test easier when related locations appear in both training and testing data. Performance dropped as the geographical test blocks became larger, showing that apparent accuracy depended on the distance separating familiar and unfamiliar locations.

Holding out an entire year exposed the clearest weakness, with average R² scores falling to 0.044 for Random Forest and −0.106 for XGBoost. Those results show very limited ability to predict prevalence in years excluded from training, despite the stronger performance across locations. The study supports further investigation of spatial mapping, without establishing a reliable system for forecasting future malaria outbreaks.

Turning promising maps into useful public health tools

Freely available satellite information offers a potentially affordable way to identify environmental patterns that could support surveillance and guide further investigation. Water-related indicators and elevation could help public health teams decide where closer monitoring deserves attention. The study did not test whether using these maps improves prevention outcomes or reduces infections.

The analysis left out socioeconomic conditions, population characteristics and malaria control measures, all of which can shape exposure and infection patterns. Bringing detailed satellite observations onto a five-kilometre grid removed local information, and the use of existing malaria estimates limits independent validation. The model's explanations describe associations within its predictions and cannot establish that a particular environmental feature causes malaria.

Future work needs independent field observations, information about prevention programmes and communities, and broader training data to assess performance across Tanzania's varied landscapes. Finer-resolution observations could help researchers examine local differences hidden by the current grid, with separate testing needed before extending the approach to other regions. The practical promise lies in combining environmental maps with evidence from communities to build a clearer picture of where malaria surveillance is most needed.

  • FIRST PUBLISHED IN:
  • Devdiscourse
Give Feedback

Use this form for editorial or site feedback. We usually reply within 2 to 3 working days.

By submitting, you agree that we may use your email address to respond.