Adaptive SPD-YOLO: Enhancing Spatial Feature Retention for Lunar Boulder Detection

Abstract: Mainstream real-time object detectors routinely sacrifice spatial resolution to keep inference costs manageable, a trade-off that proves especially damaging when the targets of interest span only a handful of pixels. This work introduces AdaptiveSPD-YOLO, a modified YOLOv26 architecture that counters this loss through an Adaptive Space-to-Depth downsampling module inserted at the P3/8 backbone junction. Rather than indiscriminately rearranging every channel into the depth dimension, the module employs a Squeeze-and-Excitation–style channel-attention gate that scores each feature map by its informational salience before the spatial-to-channel rearrangement takes place. To quantify the benefits of this selective preservation strategy, a variance-based Spatial Retention Tracking protocol is introduced and monitored across training epochs. Experiments on a large-scale lunar boulder dataset comprising 23,154 multiple scale orbital images with 8,94,474 annotated bounding boxes yield a peak mAP@50 of 78.1% and a precision of 76.8% at a computational cost of 70.6 GFLOPs. Ablation analysis confirms that the attention-gated variant surpasses both standard strided convolution and uniform SPD, while the channel-attention gate autonomously increases its suppression rate from 49.9% to 61.8% during training, indicating an emergent capacity for discriminative feature selection that directly correlates with improved detection accuracy.

Keywords: Adaptive Space-to-Depth, YOLOv26, Channel Attention, Spatial Information Retention, Lunar Boulder Detection, Small Object Detection