Zhou Ziyang, Zhang Huichun, Bian Liming, Zhou Lei, Ge Yufeng
College of Mechanical and Electronic Engineering, Nanjing Forestry University, Nanjing, Jiangsu, China.
Jiangsu Co-Innovation Center of Efficient Processing and Utilization of Forest Resources, Nanjing Forestry University, Nanjing, Jiangsu, China.
Plant Biotechnol J. 2025 Jul;23(7):2464-2481. doi: 10.1111/pbi.70039. Epub 2025 Mar 30.
Increased drought frequency and severity in a warming climate threaten the health and stability of forest ecosystems, influencing the structure and functioning of forests while having far-reaching implications for global carbon storage and climate regulation. To effectively address the challenges posed by drought, it is imperative to monitor and assess the degree of drought stress in trees in a timely and accurate manner. In this study, a gradient drought stress experiment was conducted with poplar as the research object, and multimodal data were collected for subsequent analysis. A machine learning-based poplar drought monitoring model was constructed, thereby enabling the monitoring of drought severity and duration in poplar trees. Four data processing methods, namely data decomposition, data layer fusion, feature layer fusion and decision layer fusion, were employed to comprehensively evaluate poplar drought monitoring. Additionally, the potential of new phenotypic features obtained by different data processing methods for poplar drought monitoring was discussed. The results demonstrate that the optimal machine learning poplar drought monitoring model, constructed under feature layer fusion, exhibits the best performance, with average accuracy, average precision, average recall and average F1 score reaching 0.85, 0.86, 0.85 and 0.85, respectively. Conversely, the novel phenotypic features derived through data decomposition and data layer fusion methods as supplementary features did not further augment the model precision. This indicates that the feature layer fusion approach has clear advantages in drought monitoring. This research offers a robust theoretical foundation and practical guidance for future tree health monitoring and drought response assessment.
在气候变暖的情况下，干旱频率和严重程度的增加威胁着森林生态系统的健康和稳定性，影响着森林的结构和功能，同时对全球碳储存和气候调节具有深远影响。为了有效应对干旱带来的挑战，必须及时、准确地监测和评估树木的干旱胁迫程度。在本研究中，以杨树为研究对象进行了梯度干旱胁迫实验，并收集了多模态数据以供后续分析。构建了基于机器学习的杨树干旱监测模型，从而能够监测杨树的干旱严重程度和持续时间。采用了数据分解、数据层融合、特征层融合和决策层融合四种数据处理方法对杨树干旱监测进行综合评估。此外，还讨论了不同数据处理方法获得的新表型特征在杨树干旱监测中的潜力。结果表明，在特征层融合下构建的最优机器学习杨树干旱监测模型表现最佳，平均准确率、平均精确率、平均召回率和平均F1分数分别达到0.85、0.86、0.85和0.85。相反，通过数据分解和数据层融合方法获得的新表型特征作为补充特征并没有进一步提高模型精度。这表明特征层融合方法在干旱监测中具有明显优势。本研究为未来树木健康监测和干旱响应评估提供了坚实的理论基础和实践指导。
