The efficiency of transfer learning for detecting drones using thermal imagery is highly dependent on the quality and diversity of the underlying training datasets. This chapter presents a comprehensive and reproducible pipeline for constructing high-quality aerial thermal datasets optimized for transfer learning applications. We detail the end-to-end process, encompassing thermal image acquisition, precise annotation, domain-specific augmentation, and dataset refinement through advanced cleaning techniques. To ensure robust model transferability across real-world scenarios, we emphasize the inclusion of environmental diversity—spanning diurnal cycles, weather variability, and heterogeneous landscapes such as urban and rural areas. We introduce novel thermal-specific augmentation strategies, including thermal blur, noise injection, and dynamic brightness modulation, to bridge the domain gap between pretraining and target deployment. Additionally, cleaning procedures using perceptual hashing and statistical outlier detection are employed to maintain dataset integrity and eliminate redundancy. Experimental evaluations confirm that datasets processed with our methodology yield superior performance in drone detection tasks, significantly enhancing accuracy, robustness, and generalization across varying conditions. This chapter thus provides practical insights and guidelines for researchers and practitioners aiming to leverage transfer learning in thermal imaging contexts, contributing to more reliable and scalable drone detection systems for ground-to-air surveillance.