2025/05/09 by Mahya Parchami, Negin Nazarian, Melissa Hart +2 · 1 voice · 1 citation
Engineering · #IoT-based Smart Home Systems
paper · doi:10.1088/1748-9326/add6b5
openalex publication_date 2025/05/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/22
Abstract Despite ongoing efforts to collect high-resolution datasets that capture the spatial distribution of urban heat, there remains a gap in human-centric monitoring that focuses on the immediate environment of individuals experiencing heat exposure. To address this, we explored three different models for predicting air temperature in dynamic outdoor settings using wrist-mounted wearable sensors. Data was collected for 22 d between 2020 and 2024 in Sydney, Australia. Each experiment involved 6–15 participants walking through different built environments. When air temperature and relative humidity measured by wrist-mounted sensors were compared to reference sensors, we found that wrist-mounted wearables cannot directly measure air temperature due to the influence of skin temperature. However, we can use their data to train a prediction model for air temperature. We explored three prediction methods: a steady-state heat transfer model of human skin, multi-linear regression, and random forest machine learning (ML). Results showed that the heat transfer model relied heavily on climatic parameters which could not be measured by wrist-mounted sensors, limiting the applicability of this method. The linear regression model (developed solely based on wrist-mounted data) neglected the non-linear correlation between wrist air temperature and wrist skin temperature. The ML approach, however, was capable of capturing non-linear, multi-dimensional relationships and demonstrated the best predictive performance. ML tested on out-of-sample data achieved a correlation coefficient ( R 2 ) of 0.97 (in contrast with 0.60 and 0.88 for heat transfer and linear regression) between predicted and observed air temperature, with mean absolute error of <1 °C (in contrast with 4.43 and 1.81 °C). This performance is equivalent to the accuracy of many common air temperature sensors. This prediction model can be an effective method for providing high-resolution air temperature data in cities with temperate climates, such as Sydney, while informing future work in other climate backgrounds.