Instance Segmentation and Localization of Overlapping Subsurface Rocks in GPR Images Using ConvNeXt and Hybrid Task Cascade

Small subsurface rocks pose significant challenges to drilling operations during lunar regolith sampling missions. Ground-penetrating radar (GPR), a non-destructive detection technique, provides high-resolution subsurface imaging; however, accurately interpreting rock echoes remains challenging due to signal interference. This study proposes a novel instance segmentation framework to accurately identify echo features of small rocks in GPR images. Specifically, the ConvNeXt architecture is adopted as the backbone network to enhance the extraction of global contextual features. Furthermore, an optimized Hybrid Task Cascade head is introduced to strengthen joint learning across multi-stage detection and segmentation tasks, substantially boosting model performance in complex scenarios with overlapping hyperbolic signatures. A hybrid localization strategy that integrates the least-squares hyperbolic fitting with a fully connected neural network is developed to achieve precise regression of rock spatial positions. Extensive experiments on both synthetic and field datasets demonstrate that the proposed algorithm achieves a precision of 94.3% and a recall of 94.8%, with horizontal localization errors typically below 2.5 cm and depth errors generally under 5%. The methodology offers valuable technical support for optimizing drilling path planning in future lunar exploration missions.