Abstract: Global climate change and intensified human activities have accelerated the loss of plant diversity. Traditional plant surveys, which rely on taxonomists for on-site identification, face challenges such as low efficiency, high costs, and limited coverage, making it difficult to meet the demands for large-scale, high-frequency, and detailed biodiversity monitoring. This study focuses on Wuyishan National Park as the research area. Based on the mature framework of the FlowerMate 2.0 model and integrating transfer learning with the EfficientNetV2S lightweight architecture, a localized image recognition model was developed, covering 2554 species of higher plants in this region. Model performance evaluation results show that the Wuyishan localized model achieved recognition accuracy rates of 77.9%, 86.6%, and 90.0% at the species, genus, and family levels, respectively, with precision rates of 92.1%, 94.0%, and 95.5% at these same taxonomic levels. These indicators outperform those of the national 20K model of FlowerMate Professional Edition. Furthermore, the training efficiency of the localized model is significantly higher than that of the national model. This study not only helps overcome the technical and cost bottlenecks associated with intelligent monitoring in grassroots nature reserves, but also provides a practical technological paradigm for the scientific investigation and precise conservation of plant biodiversity across different regions within national parks. It establishes a solid technical foundation for constructing human-machine collaborative plant diversity monitoring system.