LRD-RB: The First Large-Scale Dataset With Rotated Bounding Boxes for Automated Lunar Rockfall Detection

Lunar rockfalls represent a significant geological phenomenon on the Moon's surface, characterized by boulders or rock fragments moving from higher to lower terrain, leaving distinct trails. This phenomenon is important for revealing the erosion state and activity of the lunar surface. Due to the small size and sparse distribution of lunar rockfalls in satellite imagery, manual identification is highly challenging. Deep learning-based rotated object detection methods are particularly well-suited for lunar rockfall identification, as they can accurately capture the elongated shapes and diverse orientations of rockfalls and their tracks. However, no dataset is currently available for model training. To address this issue, we present the first large-scale Lunar Rockfall Dataset with Rotated Bounding boxes (LRD-RB). The dataset comprises 11,697 image patches containing 58,298 lunar rockfall features annotated with rotated bounding boxes. LRD-RB demonstrates excellent diversity in spatial resolution, solar illumination conditions, and geographical distribution, providing comprehensive feature representations for training deep learning models. We conducted systematic evaluations of several representative rotated object detection models on LRD-RB. The dataset not only facilitates the development and testing of lunar rockfall detection algorithms but also has the potential to serve as a challenging benchmark for rotated small object detection tasks in remote sensing applications. Our dataset will be released via the following link: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/MIAODRB/LRD-RB</uri>.