Objective Traditional afforestation acceptance methods are inefficient and struggle to adapt to complex scenarios. Meanwhile, the difficulty in directly using drone imagery for AI model input has hindered the intelligentization of afforestation acceptance. This study proposes a MDIEA-YOLO detection model based on drone imagery for afforestation acceptance scenarios. The goal is to achieve efficient identification and accurate counting of afforestation seedlings, thereby improving the precision and efficiency of afforestation acceptance and providing technical support for modernization of forestry management.

Method To achieve the above objectives, this study developed a “multi-dimensional interaction enhanced attention module” (MDIEA). By integrating convolutional block attention mechanism and Shuffle Attention mechanism, MDIEA can efficiently process features of complex scenarios and small targets, significantly enhancing the network’s parsing ability. Embedding MDIEA into the YOLOv8 feature extraction network, the refined channel and spatial attention weighting strengthened recognition of key features. In addition, the XIoU loss function was introduced to optimize model’s boundary localization of small and overlapping targets, further improving detection accuracy. Ultimately, an end-to-end image preprocessing pipeline based on drone imagery and the MDIEA-YOLO model was constructed to enable automatic identification and counting of afforestation seedlings.

Result In experiments conducted at the Jiangle State Forest Farm in Fujian Province of eastern China, the MDIEA-YOLO model achieved mAP0.5 values of 97.5%, 96.1%, and 96.8% on the 1-year-old, 2-year-old, and 3-year-old datasets, respectively, significantly outperforming other comparative models. Under different lighting and resolution conditions, the MDIEA-YOLO model’s mAP0.5 values remained above 92%, demonstrating good robustness. When processing 100 images, the CPU and GPU processing efficiencies of MDIEA-YOLO model were similar, indicating high flexibility and adaptability of the system in practical applications. Comparison with manual inspection revealed that the system exhibited comparable or even higher accuracy and efficiency in key indicators, proving its reliability and practicality.

Conclusion The afforestation acceptance drone image preprocessing system proposed in this study has effectively advanced the intelligentization of afforestation acceptance, significantly enhancing acceptance efficiency and accuracy. It offers a new technical solution for the afforestation acceptance field with broad application prospects. In the future, model the performance will continue to be optimized, and the dataset scale will be expanded to accommodate a wider range of application scenarios, promoting the modernization of forestry management.