Object detection is crucial in analyzing imagery from satellite sensing, with developments in geospatial monitoring technology and deep convolutional networks driving significant progress. Nevertheless, difficulties remain, especially in recognizing objects across different sizes and precisely identifying minor features. This paper presents a novel network, Customized YOLOv8, to tackle these issues and enhance object detection capabilities. Many researchers use digital elevation models (DEMs) for crater classification, but their low resolution often limits the detection of smaller craters. In response, this paper leverages lunar CCD images from NASA’s 100-m global mosaic, captured by the Lunar Reconnaissance Orbiter Camera (LROC) WAC. The Customized Yolov8 model is built with Dense-Net121 as the backbone for efficient feature reuse, improving gradient flow and minimizing vanishing gradients. This paper used Bi-FPN as the neck for fusing multi-scale features, which is essential for recognizing objects of various dimensions. Additionally, Soft-NMS enhances accuracy by reducing missed detections in closely overlapping objects. This architecture achieves 76.0% precision, 68.6% recall, and a minimum average precision of 78.7% on test data without any prior preprocessing applied to the raw image.