Abstract: Australian nut flowers, known as macadamia inflorescences, is one variety of the Australian native trees in Yunnan Province, China. However, large-scale cultivation under high altitude has led to frequent off-season flowering and asynchronous flowering in the same tree. Reducing flowering synchrony is also detrimental to pollination, fertilization, fruit set, fruit retention, and uniform harvesting, thereby affecting yield stability and fruit quality. In addition, conventional flowering regulation can rely heavily on manual operations, leading to labor-intensive, low efficient, costly, and highly dependent on subjective experience. As a result, it is urgent demand for the high precision regulation of the macadamia inflorescences in smart orchard. In this study, an intelligent identification was proposed for macadamia inflorescences under natural environments using an improved YOLO11n object detection model. Specifically, several challenges were solved in real orchard environments, including complex background interference, multi-scale morphologies of inflorescences, partial occlusion, and the constraints of real-time edge deployment. 1) Partial Multi-Scale Feature Aggregation (PMSFA) module was introduced to integrate the partial convolution with multi-scale feature extraction. Multi-source heterogeneous features were efficiently captured from different network layers. Thereby, the small and easily occluded inflorescences were detected under cluttered natural backgrounds. 2) A Slim-neck architecture was constructed to balance accuracy and computational efficiency using Group-Shuffle Convolution (GSConv) and the Variety of View Group Shuffle Cross-Stage Partial Network (VoV-GSCSP). Computational complexity and parameter size were reduced to maintain the high accuracy of the detection, thus improving the edge deployment. 3) In terms of loss function optimization, Focaler-PIoU v2 was selected to replace the original Complete Intersection over Union (CIoU) loss in YOLO11n. The stability of bounding-box regression was improved from the gradient contribution of high-quality samples during training. Localization accuracy was enhanced under challenging conditions, such as occlusion and scale variation. A comparison was finally conducted on the macadamia inflorescences dataset from Yunnan Province. The results demonstrated that the improved YOLO11n model consistently outperformed mainstream object detection over all evaluation metrics. Specifically, precision increased by 20.1, 5.4, 1.7, 3.5, 3.6, 2.0, and 0.8 percentage points, respectively, while Recall increased by 1.4, 16.2, 8.1, 5.0, 6.4, 9.1, and 6.7 percentage points, respectively, compared with the Faster R-CNN, SSD, RT-DETR, YOLOv5s, YOLOv8n, YOLOv10n, and YOLOv12n. Mean average precision at an IoU threshold of 0.5 (mAP_0.5) was also improved by 6.4, 15.9, 5.7, 4.3, 4.1, 5.0, and 3.3 percentage points, respectively. Furthermore, the improvements reached 20.1, 33.5, 11.9, 13.5, 10.6, 11.0, and 9.7 percentage points, respectively, under the more mAP_0.5–0.95 threshold. There were superior precision, recall, and overall detection accuracy under complex natural conditions. To verify the feasibility, the improved model was also deployed on the Orange Pi AI Pro edge-computing platform. A real-time detection speed of 72.8 frames per second (FPS) was achieved using Neural Processing Unit (NPU) inference with multi-threaded scheduling, fully meeting the requirements of accurate and real-time recognition in natural orchard environments. Overall, there was a favorable balance among detection accuracy, computational efficiency, and deployment feasibility. The finding can also provide strong technical support for the precise flowering-period regulation for macadamia production in Yunnan Province, China.