Toward transferable EO-based modelling of intra-slum air temperature using citizen-science data: evidence from Nairobi and Lagos

Source abstract

Urban areas are typically warmer than their non-urban surroundings, a phenomenon known as urban heat island (UHI); yet, intra-urban temperature contrasts remain under-characterized, especially within and between slums. Conventional meteorological stations, often located at city peripheries (e.g., airports), poorly capture neighbourhood-scale microclimates. Earth Observation (EO) provides land surface temperature (LST), which is widely used to characterize urban heat islands. However, LST is not a direct proxy for near-surface air temperature (Tair), as the LST-Tair relationship is modulated by several interacting factors such as heat storage, vegetation, ventilation, and urban canopy effects among others. We investigate whether intra-slum Tair differences can be modelled using EO covariates alone and test potential transferability of such models across African cities. Quality-controlled, citizen-generated Tair observations collected between January and March 2024 are aggregated to 50 × 50 m grid cells and used as the target variable in a Random Forest (RF) model. EO-derived features (albedo and radiative properties, urban morphology indicators and vegetation index), computed on the same grid, serve as covariates for modelling intra-urban Tair to generate high-resolution heat maps within and between slums in Nairobi (Kenya) and Lagos (Nigeria). Trained in Nairobi, the model attains a low out-of-sample mean absolute error, however a direct transfer of the Nairobi-trained model to Lagos underperforms, indicating that zero-shot transferability is limited. Instead, robust performance requires a domain-adaptation strategy. Incrementally adding a modest fraction of Lagos samples to the training set improves cross-city generalisation, yielding comparable accuracy in both cities. Beyond demonstrating the value of low-cost sensors and citizen science for actionable climate information, we release an open, reproducible pipeline to guide urban-cooling interventions and benchmark resilience policies where evidence is most limited.