Abstract: Harvesting timing can directly determine the flavor quality of the fruit in large-scale intelligent cultivation. However, the uneven spatial distribution of factors, such as light and nutrients, can lead to variations in the fruit maturity among different trees, even among various parts of the same tree. Conventional detection can rely primarily on manual experience for sampling and discrimination. It is often required to accurately assess the overall maturity of the peach orchard in the sustainable industry. Furthermore, the differing ripening times of fruits in natural environments can combine with the complex interferences, such as foliage occlusion and fruit overlap. It is further difficult to accurately identify the peach maturity in orchards. According to the maturity requirements for harvested peaches in the national standard (NY/T 586-2002), and the technical code for peach storage (GB/T 26904-2020), previous study has conducted to classify the peach fruit maturity into three categories using color: 1) Unripe peach (up): The fruit surface is entirely green with no sign of color change; 2) Half-ripe peach (hp): The fruit surface can share a mix of red and green, where the color-changed area is less than 50%; 3) Ripe peach (rp): the color-changed area on the fruit surface can exceed 50%. In this study, an improved YOLOv8n model was proposed to identify the peach maturity in orchards. 1) The original Conv module was replaced with the PMSEConv (Peach Multi-Scale Efficient Convolution) module to capture multi-scale contextual features with fewer missed and false detections. 2) Peach Occlusion Attention mechanism (POAttention) was introduced into the Neck section to improve the detection accuracy for peach maturity in high-density and occluded environments. Finally, the EIoU loss function was adopted to optimize the bounding box regression, thereby enhancing the adaptability and practicality in complex orchard environments. Ablation experiment indicated that the improved YOLOv8n model shared better performance than before. Compared with the baseline YOLOv8n, the precision and recall rates increased by 5.2 and 2.8 percentage points, respectively, and the mean average precision (mAP@0.5) was improved by 5.9 percentage points. Simultaneously, the model complexity was effectively controlled, where the FLOPs and parameters were reduced by 0.5 G and 0.29 M, respectively, with a dual optimization of accuracy and computational efficiency. Compared with current mainstream object detection models, including RT-DETR, YOLOv3-tiny, YOLOv5, YOLOv6, YOLOv9, YOLOv10, YOLOv11, and YOLOv12, the improved YOLOv8n model was achieved in the optimal key metrics, with the precision of 84.5%, recall of 76.0%, and mAP@0.5 of 83.8%. Excellent performance was obtained in identifying the fruit maturity in complex orchard environments. The lightweight and efficient model can provide strong technical support for peach-picking robots in an intelligent orchard, thereby facilitating precise and efficient operations.