Vector-borne diseases, such as dengue, have reached record epidemic levels in the Americas and are emerging as a new threat in Europe. Traditional surveillance of mosquito breeding sites remains inefficient. While the use of UAVs (drones) and Deep Learning is promising, it faces a critical methodological challenge: class imbalance, which is considered severe in real-world data. This imbalance biases standard models, causing them to ignore the most epidemiologically critical vector sites (minority classes). This paper presents an experimental comparative analysis of four state-of-the-art object detection architectures (YOLOv8m, YOLOv11m, YOLOv12m, and RF-DETR) to evaluate their robustness in this scenario. Performance was measured using standard metrics (mAP) and specialized metrics (Balanced Accuracy, MCC). The results demonstrate that the baseline (YOLOv8m) fails to detect minority classes, achieving a recall of only 8.2% for the bottle class. The YOLOv11m architecture, equipped with spatial attention mechanisms (C2PSA), emerged as the optimal solution. It achieved the highest Balanced Accuracy (68.4%) and MCC (0.487), and improved the bottle class recall by 247% compared to the baseline. We conclude that architectures incorporating spatial attention mechanisms are crucial for the viability of automated epidemiological surveillance in real-world, severely imbalanced environments.