Microclimates are critical for understanding how organisms interact with their environments, influencing behaviour, physiology, and species distributions. However, traditional physical heat-balance models for predicting ground temperatures in microhabitats often exhibit biases due to unaccounted environmental complexities and poorly constrained parameters. These limitations can hinder ecological research and conservation planning, particularly in the context of climate change. In this study, we demonstrate how high-resolution drone-based mapping and machine learning can improve the accuracy of microclimate models. Using drone imagery, we generated detailed environmental maps, including solar radiation, vegetation indices, and skyview factors, to parameterize a physical heat-balance model. Validation with thermal maps derived from drone-mounted infrared cameras revealed systematic errors in the physical model's predictions, including over- and underestimations under specific environmental conditions. To address these errors, we applied a random forest machine learning model to predict and correct biases in new prediction maps. Our results show that machine learning reduced mean absolute errors by over 30% and mean square errors by 50%, while consistently narrowing the range of prediction inaccuracies. Key factors driving biases, such as vegetation cover, solar radiation, and height above ground, were identified, offering valuable insights for improving physical models. The machine learning corrections not only improved accuracy but also highlighted parameters and processes that were previously underrepresented or oversimplified in traditional models. These findings illustrate the potential of machine learning to improve microclimate predictions. While our drone-based approach is most applicable to open, sparsely vegetated habitats, the principle of machine learning bias correction can be extended to other systems as well. Correcting microclimate models with machine learning and observational data provides ecologists and conservation practitioners with a powerful framework for generating more accurate microclimate estimates. Such improvements deepen our understanding of species' responses to climate change and support climate-resilient management strategies.

