In high-altitude complex terrain environments, exploration engineering targets such as trenches and drilling platforms exhibit characteristics of feature blurring, strong background interference, and extremely small scales in remote sensing imagery, resulting in significant performance degradation of existing object detection algorithms. To address these challenges, this study proposes an improved YOLOv8 object detection algorithm integrating shallow feature enhancement and scale-aware dynamic weighting from three synergistic dimensions: feature enhancement, detection optimization, and loss improvement. First, an Enhanced feature extraction Module with Attention (EMA) is designed, which introduces a bottom-up enhancement branch in the Path Aggregation Network (PANet) and embeds a lightweight spatial attention unit based on parallel dilated convolutions to alleviate small object feature dilution. Second, the detection head is optimized by adding a P2 super-shallow detection branch and customizing anchor prior parameters through K-means clustering. Third, a Scale-aware Weighted Loss (SWL) function based on the inverse of target area is proposed to balance optimization contributions across different scales. Experimental results on a self-built plateau exploration engineering dataset (12,795 samples) demonstrate that the proposed method achieves mAP@0.5 of 71.6% and mAP@0.5:0.95 of 68.5%, surpassing the YOLOv8n baseline by 4.4 and 3.8 percentage points, respectively. Precision and Recall reach 74.5% and 66.3% with an inference speed of 112 FPS at 3.2M parameters. Ablation experiments confirm independent contributions of each module: EMA improves mAP@0.5 by 1.9%, P2 branch by 1.3%, and SWL by 1.7%. After structured pruning and quantization, the model is compressed to 3.24M parameters with 215 FPS inference speed, incurring only 1.5 percentage points mAP loss.