This paper introduces the Focal Hanning Loss, a novel approach to improve UAV self-localization using heatmap classification. Building upon recent advancements in vision-based localization methods, we address limitations in existing techniques, particularly the Hanning Loss used in WAMF-FPI. Focal Hanning Loss incorporates a modulation factor that dynamically adjusts the importance of samples based on their difficulty, effectively mitigating issues such as red clusters in heatmaps and imbalanced weighting between positive and negative samples. Additionally, we enhance the model’s robustness through data augmentation techniques, which include random cropping and scaling of satellite images to simulate multi-scale variations. To validate the effectiveness of Focal Hanning Loss combined with data augmentation, we trained the WAMF-FPI model on a custom dataset derived from DenseUAV, achieving a 3.4-point improvement in the Relative Distance Score metric compared to original Hanning Loss.