A geospatial modeling study conducted across Dar es Salaam, Tanzania, has demonstrated that combining three-dimensional convolutional neural networks, long short-term memory recurrent units, and cellular automata can predict local malaria transmission patterns with high precision while forecasting long-term environmental risk dynamics through the year 2060. The research, published in the International Journal of Health Geographics by investigators Edmund Steven Kanjagaile, Dorothea Deus, and Anastazia Daniel Msusa, addresses persistent public health challenges in tropical coastal cities where Anopheles mosquito breeding habitats remain difficult to track through manual ground surveys alone. Conventional entomological field inspections across rapidly expanding urban centers are often constrained by high labor costs, fragmented physical access, and incomplete reporting. To overcome these logistical barriers, the research team constructed an integrated computational architecture that fuses multi-source satellite imagery, meteorological observations, demographic records, and routine clinical notifications into a unified spatiotemporal surveillance platform.
The study was set in the coastal metropolis of Dar es Salaam, a low-lying region spanning latitudes 6.50 degrees to 7.80 degrees South and longitudes 39.00 degrees to 39.52 degrees East, encompassing five administrative districts: Kinondoni, Ubungo, Ilala, Temeke, and Kigamboni. According to data from the 2022 National Bureau of Statistics census, the commercial capital accommodates 5,383,728 residents living in 1,550,066 households across varied socio-demographic and commercial settings. The coastal city exhibits an average annual temperature of 27.3 degrees Celsius, which is approximately 3.08 percent higher than Tanzania’s national average, alongside an average relative humidity of 77 percent. Precipitation totals roughly 1,148 millimeters annually across approximately 224 rainy days, with the most intense rainfall occurring between mid-March and mid-May and the driest conditions settling between July and September. Uneven terrain and drainage deficits make the metropolis highly susceptible to recurrent urban flooding, creating extensive standing water collections that support vector proliferation in close proximity to human dwellings.
To evaluate the environmental foundations of transmission, the researchers assembled a multifaceted spatial dataset spanning from 2015 through 2023. Facility-level epidemiological observations comprised clinical malaria records gathered across 1,794 health facilities reporting through the national District Health Information Software 2 platform, accessed via the Africa Data Catalog. Gridded monthly meteorological layers detailing precipitation, temperature, and specific humidity at a four-kilometer spatial resolution were acquired from the TerraClimate repository and calibrated against Tanzania Meteorological Authority station records situated at DAR-PORT. Remotely sensed spectral surface parameters were obtained at thirty-meter spatial resolution from Landsat 7, Landsat 8, and Landsat 9 platforms across three distinct temporal benchmarks in 2015, 2019, and 2023. Bilinear interpolation and inverse distance weighting were executed to resample diverse climate and case records to a standardized analytical grid, while decennial census records from 2012 and 2022 were projected at an annual national population growth rate of three percent and mapped to continuous thirty-meter raster surfaces.
Before deep learning execution, the team engineered land surface metrics to capture ecological conditions governing the biological lifecycle of Plasmodium falciparum and its principal mosquito vectors, including the Anopheles gambiae sensu lato species complex and Anopheles funestus sensu lato. Land use and land cover were categorized into five ecologically defined classes comprising open water, built-up surfaces, vegetation cover, swamps, and bare land using a Random Forest classifier evaluated on an eighty-twenty training and testing split. The Normalized Difference Vegetation Index was generated from red and near-infrared reflectance bands to quantify vegetative moisture and density, while the Normalized Difference Water Index was derived to trace water body dynamics. All seven predictor channels were normalized to a zero-to-one range through feature scaling, and a Random Forest ensemble quantified predictor contributions. Climate metrics emerged as the primary operational driver, representing 56.02 percent of overall importance through temperature at 20.37 percent, humidity at 18.56 percent, and rainfall at 17.09 percent. Environmental indices contributed 26.65 percent, while population density accounted for 17.33 percent.
The computational pipeline employed a deep neural network architecture configured to decouple local spatial neighborhood relationships from long-term temporal dependencies across the 2015, 2019, and 2023 observation steps. Input arrays were structured into five-dimensional tensors measuring 1,948 samples by 68 rows by 61 columns across three temporal intervals and seven feature channels. A three-dimensional convolutional neural network formed the initial structural layer, employing thirty-two and sixty-four filters with a kernel dimension of three by three by two. This three-by-three spatial window spanned a ninety-meter by ninety-meter land surface at thirty-meter resolution to preserve intra-ward heterogeneity, while the temporal kernel dimension of two encompassed contiguous four-year transitions. Three-dimensional max-pooling operations reduced spatial dimensions while Leaky ReLU activations prevented vanishing gradients. The resulting spatial feature representations were subsequently fed into a two-layer Long Short-Term Memory network utilizing sixty-four and thirty-two hidden units with hyperbolic tangent activations, followed by fully connected dense layers and a thirty percent dropout regularization layer to prevent structural overfitting.
To simulate how localized transmission risk spreads physically across adjacent neighborhoods, the model integrated a Cellular Automata module over a three-by-three Moore neighborhood. The probability of risk escalation within each analytical cell was dictated by a logistic sigmoid transition function parameterized by an inflection threshold gamma calibrated at 0.293 and a transition steepness parameter beta set at 43.9. These parameters were derived empirically from the 2015 to 2023 baseline data, where the inflection threshold reflected the mean baseline risk separating stable cells from evolving zones, and the steepness parameter was extracted from the ninetieth percentile of risk increments constrained to yield an approximate two percent growth across five-year iteration steps. When local neighborhood conditions surpassed the calibrated threshold of 0.105, the cell risk was adjusted upward by an increment value of 0.0834, capped at the theoretical maximum boundary. Incorporating the Cellular Automata module raised the spatial clustering of risk, shifting Moran’s I spatial autocorrelation from 0.72 in the raw convolutional network output to 0.81 in the final hybrid prediction surface.
Model training carried out over twenty-five epochs with a batch size of thirty-two and a validation split of twenty percent demonstrated marked convergence, reducing training loss by 97.30 percent and validation loss by 90.15 percent. Evaluated against a naive mean-predictor baseline mean squared error of 0.1, the fully integrated hybrid framework achieved an overall test mean squared error of 0.005, denoting a ninety-five percent reduction in predictive error. When validated against observed clinical malaria incidence across Dar es Salaam health facilities in 2023, the spatial predictions demonstrated strong statistical fidelity, registering a Spearman rank correlation coefficient of 0.92 with a significance value of p below 2.2e-16. Facility-level case distributions corroborated the model’s spatial stratifications, as health centers located within wards designated as high risk reported significantly greater patient loads than facilities in medium-risk wards under Kruskal-Wallis testing (H equals 9.46, p equals 0.002), an outcome confirmed among active reporting centers by a one-sided Mann-Whitney U test (U equals 2,776, p equals 0.035).
The spatial outputs delineated distinct risk zones across the urban landscape, classifying normalized risk scores into low risk from 0.00 to 0.33, medium risk from 0.34 to 0.66, and high risk from 0.67 to 1.00. While no wards reached full saturation at the 1.00 ceiling, high risk concentrated heavily along southern peri-urban boundaries. Wards identified with high transmission intensity included Pembamnazini, Tundwi, Msongola, Zingiziwa, Chamanzi, Kivule, and Mbweni. Clinical records reflected these findings: during 2023, Tundwi Songani Dispensary logged 1,489 confirmed malaria cases, Zingiziwa Health Center recorded 450 cases, two dispensaries in Kivule jointly reported 440 cases, Msongola Dispensary documented 213 cases, and Mbweni Dispensary treated 210 cases. Conversely, centrally located wards within Kinondoni and Ilala exhibited lower detected transmission risk despite maintaining high commercial and residential densities, reflecting reduced indoor vector biting and the displacement of larval breeding habitats by contiguous built-up surfaces.
A critical observation highlighted by the researchers involves the inverse relationship between healthcare facility infrastructure and underlying transmission vulnerability. Densely populated central urban areas house abundant health facilities but demonstrate minimal malaria risk, whereas peripheral wards such as Kigamboni, Msongola, and Tundwi exhibit elevated epidemiological risk while being served by relatively few diagnostic centers. Longitudinal analysis of total citywide clinical notifications between 2015 and 2023 indicated a declining trend across reporting centers, with paired facility evaluations showing statistically significant reductions under the Wilcoxon signed-rank test across 813 facilities (p below 0.001). The authors attribute this historical reduction to expanded vector control interventions, including long-lasting insecticidal nets, targeted indoor residual spraying, and improved clinical case management, which tempered transmission despite ongoing demographic expansion.
Looking toward future climate trajectories, the research team projected malaria risk surfaces at five-year intervals from 2025 through 2060 by iteratively propagating the calibrated Cellular Automata rules across Dar es Salaam. The projections reveal a steady, gradual spatial expansion of elevated transmission zones without abrupt systemic disruptions. Across the thirty-five-year horizon, maximum projected risk climbed from 0.763 in 2025 to 0.903 in 2060, remaining entirely within the high-risk band without reaching the theoretical ceiling of 1.00. The mean risk difference across the full projection interval was 4.14 percentage points, advancing along an ordinary least squares linear trajectory with an annual slope of 0.0022 risk units per year, equivalent to a cumulative risk increase of 0.077 units or 7.7 percent by 2060. Five-year growth rates displayed slight non-monotonic fluctuations, declining from 3.26 percent in 2030 to 3.15 percent in 2035, edging up to 3.36 percent in 2040, and settling back to 3.15 percent in 2050.
Spatially, the 2060 projections indicate that climate-driven risk intensification will impact districts unevenly. More than three-quarters of the southern Kigamboni district is projected to experience intensified risk, followed by peripheral portions of Ilala, while the northern district of Kinondoni remains largely stable due to extensive urban infrastructure and limited microclimatic vector receptivity. However, the study authors emphasize several methodological limitations inherent in long-range predictive modeling. The Cellular Automata projections assumed stationarity in environmental-disease relationships, holding transition parameters fixed across thirty-five years. Biological dynamics such as evolving vector resistance to chemical insecticides and parasite tolerance to artemisinin were omitted from the neural feature tensors. Furthermore, demographic projections relied on a uniform three percent annual growth rate that omitted ward-level migration differences, and climate data downscaling from four-kilometer grids to thirty-meter surfaces lacked localized microclimate variations.
Despite these documented constraints, the authors conclude that the hybrid deep learning framework provides actionable spatial intelligence for municipal health authorities. Because the projected cumulative risk growth of 7.7 percent represents a modest, incremental trajectory driven predominantly by climate and population pressures, well-timed public health adaptations have substantial capacity to disrupt future transmission pathways. Expanding routine diagnostic and clinical infrastructure across peri-urban margins, implementing targeted larval source management in identified peripheral breeding reservoirs, and incorporating ecological surveillance into ongoing urban development frameworks could counteract anticipated risk increases, directly reinforcing national vector control strategies and sustainable disease prevention goals across sub-Saharan Africa.
Subject of Research: Spatiotemporal surveillance and deep learning projection of urban malaria risk in Dar es Salaam
Article Title: Geospatial-based surveillance of malaria risk in Dar es Salaam using a hybrid 3DCNN + LSTM and CA model
Article References: Kanjagaile, E. S., Deus, D., & Msusa, A. D. (2026). Geospatial-based surveillance of malaria risk in Dar es Salaam using a hybrid 3DCNN + LSTM and CA model. International Journal of Health Geographics, 25(1), Article 47. https://doi.org/10.1186/s12942-026-00482-6
Image Credits: AI Generated
DOI: 10.1186/s12942-026-00482-6
Keywords: Malaria surveillance, Dar es Salaam, Deep learning, 3D Convolutional Neural Network, Long Short-Term Memory, Cellular Automata, Random Forest, Geospatial modeling, Remote sensing, TerraClimate, Plasmodium falciparum, Anopheles gambiae, Public health informatics, Spatiotemporal analysis, Epidemiological forecasting, Urban malaria, Tanzania health data, Disease ecology
News Source: Blake Davidson. (November 19, 2026). Deep Learning Hybrid Framework Maps Dar es Salaam Malaria Risk to 2060. Scienmag.



