COMPARATIVE ANALYSIS OF YOLOv8 AND YOLOv11 FOR COLD SPOT DETECTION ON THE LUNAR SURFACE

Lunar cold spots are thermal anomalies associated with fresh impact craters and understanding them offers critical insights into the Moon's surface evolution and thermophysical properties. Traditionally, their detection has relied on manual methods, which are labor-intensive and time-consuming. This study evaluates the performance of two advanced deep learning-based object detection models, YOLOv8 and YOLOv11, for automating lunar cold spot detection using Diviner radiometer data. The training dataset was generated from 128-pixel-per-degree (ppd) rock-free nighttime regolith temperature maps covering latitudes up to ±60°. The dataset included 384 lunar images with 652 annotated cold spots for model training. For testing, the 2023 High-Resolution Nighttime Temperature dataset was cropped into 512×512-pixel sub-images (~4×4 degrees) with a 20% overlap to capture edge cold spots. This process generated 4,816 sub-images, ensuring comprehensive coverage and minimizing missed detections.

The experimental design included two strategies: a straightforward train-test split and a more robust 5-fold cross-validation approach. The models were assessed using key performance metrics: precision, recall, F1 score, and mean Average Precision (mAP). YOLOv11 consistently outperformed YOLOv8 across most metrics, achieving a precision of 0.85, recall of 0.78, F1 score of 0.81, and mAP-50 of 0.79 with K-fold cross-validation. Both models demonstrated superior performance in detecting faint thermal anomalies, showcasing their capability to identify subtle features often overlooked by manual methods.

Hyperparameter tuning and robust preprocessing techniques, including overlapping sub-image and data augmentation, contributed significantly to the models' performance. YOLOv11's higher selectivity resulted in fewer false positives and greater reliability, whereas YOLOv8 identified a larger number of cold spots, though with a higher false positive rate. Both models significantly outperformed manual detection methods, demonstrating their ability to expand the catalog of lunar cold spots efficiently and accurately with precision of 78% and 89% for YOLOv8 and Yolov11, respectively. This automated approach identified previously undetected cold spots, providing a more comprehensive understanding of lunar thermal anomalies and their spatial distribution.

These findings highlight the transformative potential of convolutional neural networks (CNNs) in planetary science, particularly in automating complex and data-intensive tasks like lunar cold spot detection. The scalability and precision of YOLOv11, combined with YOLOv8's sensitivity to faint anomalies, underscore the value of integrating deep learning techniques into planetary exploration and research.