Transferability of Machine Learning‐Based Calibration Models for Low‐Cost <scp>PM</scp> _2.5 Sensors Across Heterogeneous African Cities

ABSTRACT The expansion of low‐cost particulate matter (PM 2.5 ) sensor networks offers an opportunity to strengthen air quality monitoring and management across Africa, where reference‐grade instrumentation remains sparse. However, their effectiveness depends on robust calibration against reference‐grade monitors. Machine learning (ML) calibration methods have proven effective in correcting low‐cost sensor biases under local collocation conditions, but their scalability and transferability across heterogeneous environments have not been systematically evaluated. As sensor networks expand beyond well‐instrumented cities, calibration is no longer solely a modelling challenge but a deployment and scalability problem. This study assesses the transferability of ML‐based calibration models across five monitoring sites in four African cities representing diverse meteorological patterns and PM 2.5 concentration profiles. We compare four calibration strategies: (1) localised calibration using on‐site reference monitors; (2) direct transfer without local reference data; (3) multicity pooled models evaluated using leave‐one‐out cross‐validation (LOOCV) and (4) environmental clustering, in which cities were grouped using K‐means according to similarity in meteorological conditions and PM 2.5 distributions prior to LOOCV evaluation (cluster‐based LOOCV). Results show that localised calibration consistently provides the highest predictive performance, confirming the value of site‐specific reference data. Direct transfer performs poorly when applied to cities not from the same cluster, highlighting the risks of naive model transfer. In contrast, both LOOCV and cluster‐based LOOCV improve ML model generalisation, with cluster‐based LOOCV achieving performance comparable to LOOCV despite requiring less training data. These findings indicate that environmental similarity is a key determinant of calibration transferability. Our results demonstrate that strategically placed reference‐grade monitors in representative environmental clusters can support scalable calibration of distributed low‐cost networks across multiple cities. This approach reduces reliance on extensive collocation campaigns while maintaining acceptable performance. The findings provide a practical framework for designing scalable, data‐driven air quality networks in resource‐constrained settings.