Fully automated crater detection based on the tree-structured parzen estimator sampler technique

To comprehend the planetary body's evolution, efficient recognition of craters plays a vital role. Initially, crater detection relied on manual observation, which is prone to 40 % disagreement, varying from one researcher to another. Then, various researchers developed techniques for all object detectors based on deep learning (DL) that efficiently detected multi-scale craters. However, researchers have employed hybrid object detectors to perform multi-scale crater identification on the planetary surface by combining the strengths of multiple frameworks, such as a detection head and a feature extractor. As a result, these models suffered from suboptimal hyperparameter settings without optimisation, leading to overfitting, longer training times or reduced detection accuracy. To further refine the hybrid detector's performance, this study applies the Tree-Structured Parzen Estimator (TPE) optimisation technique to automatically tune the critical hyperparameters by identifying the best combinations of learning rate, anchor ratios, batch size, dropout, and other key parameters. On the crater dataset, the suggested model achieves a [email protected] of nearly 92.0 %, a recall of 90.2 %, and a precision of 94.8 %. These findings indicate the effectiveness of the optimisation strategy in accurately detecting craters under diverse terrain conditions, such as occluded, noisy regions with small-scale detections ∼< 0.1 km. The proposed framework of this study demonstrates strong potential for future integration into planetary mapping missions and real-time lunar exploration systems, where both accuracy and computational efficiency are critical.