高级检索
    黄成威, 齐磊, 多杰才仁, 张怀清, 薛联凤, 云挺. 基于级联循环网络的林木生长参数预测[J]. 北京林业大学学报, 2023, 45(8): 94-108. DOI: 10.12171/j.1000-1522.20230027
    引用本文: 黄成威, 齐磊, 多杰才仁, 张怀清, 薛联凤, 云挺. 基于级联循环网络的林木生长参数预测[J]. 北京林业大学学报, 2023, 45(8): 94-108. DOI: 10.12171/j.1000-1522.20230027
    Huang Chengwei, Qi Lei, Duojiecairen, Zhang Huaiqing, Xue Lianfeng, Yun Ting. Prediction of tree growth parameters based on cascaded recurrent network[J]. Journal of Beijing Forestry University, 2023, 45(8): 94-108. DOI: 10.12171/j.1000-1522.20230027
    Citation: Huang Chengwei, Qi Lei, Duojiecairen, Zhang Huaiqing, Xue Lianfeng, Yun Ting. Prediction of tree growth parameters based on cascaded recurrent network[J]. Journal of Beijing Forestry University, 2023, 45(8): 94-108. DOI: 10.12171/j.1000-1522.20230027

    基于级联循环网络的林木生长参数预测

    Prediction of tree growth parameters based on cascaded recurrent network

    • 摘要:
        目的  树木的生长参数在林木性状评估、森林碳汇计量和优化森林经营策略等方面具有重要的意义。从激光雷达数据中精准提取林木生长参数并对树木未来生长参数进行预测,以期为林业数字化发展提供技术支持。
        方法  本研究提出了一种基于人工智能的林木生长参数提取与预测方法,该方法以南京林业大学中的樱花、银杏、鹅掌楸、水杉4个树种为实验对象。首先,采用机载激光雷达获取4个树种样地的点云数据,并通过单株分割算法提取单棵树木点云。其次,基于圆拟合及高斯滤波的方法自动的从2016、2018、2020年的单棵树点云中提取胸径、树高、冠宽等生长参数,并辅以样地调查数据(2015、2017、2019、2021、2022年),构建不同树种的单棵树生长参数时间序列作为深度学习网络的训练样本集。最后,构造由两层门控循环单元(GRU)的林木生长参数预测深度学习网络,并引入注意力模块以弥补传统循环神经网络捕获长期依赖关系的不足。该网络以2015—2021年单株树木生长参数时间序列作为输入,依托训练数据及随机梯度下降算法使网络参数逼近真实树木生长情况,并用以预测2022年单棵树木的生长参数。
        结果  深度学习网络在树高预测上表现最好,决定系数R2均不低于0.83,均方根误差(ERMS)均小于0.50 m,在4种树中银杏树的预测结果最优(R2 = 0.95,ERMS = 0.31 m)。在胸径、冠宽等参数的预测上,深度预测网络仍有着良好的表现,R2均不低于0.81,胸径ERMS小于2.50 cm,冠宽ERMS小于0.32 m。在与线性回归和LSTM网络等林木参数预测方法的比较中,本模型预测效果良好(R2 ≥ 0.86),误差较小。
        结论  级联循环神经网络可有效地预测未来树木的生长情况,提高林木生长参数的预测精度,同时GRU和注意力机制的引入在林木参数的时序预测中有一定的鲁棒性,为森林的智能管理与可视化分析提供了新的思路。

       

      Abstract:
        Objective  Accurate acquisition of tree growth parameters is a necessary prerequisite for obtaining forest information and phenotypic characteristics and is of great significance in forest character assessment, forest carbon sink measurement, and optimization of forest management strategies. Nowadays, accurate extraction of tree growth parameters from light detection and ranging data and prediction of tree growth parameters in the future will provide technical support for the digital development of forestry.
        Method  In this study, an artificial intelligence based approach to tree growth parameter extraction and prediction was proposed. Four tree species in Nanjing Forestry University, i.e., sakura, ginkgo, liriodendron, and Chinese fir, were selected as experimental objects. First, airborne laser scanning was used to obtain the point cloud data of four tree species sample plots, and the individual tree segmentation algorithm was used to obtain a single tree point cloud. Second, growth parameters such as DBH height, tree height, and crown width were automatically extracted from the single tree point cloud using the circle fitting and Gaussian filter algorithms, and supplemented with artificial in-situ measurement data (measured in 2015, 2017, 2019, 2021, and 2022). The time series of single tree growth parameters of different tree species were constructed as the training sample set for the deep learning network. Finally, a deep learning network for predicting tree growth parameters was constructed with a two-layer Gated Recurrent Unit (GRU) network, and an attention module was introduced to compensate for the shortcomings of traditional recurrent neural networks in capturing long-term dependencies. The network took the time series of individual tree growth parameters from 2015 to 2021 as the input, relied on training data and stochastic gradient descent algorithm to approximate the network parameters to the real tree growth situation, and was used to predict the growth parameters of individual trees in 2022.
        Result  The network performed best in the prediction of tree height, with the R2 not less than 0.83 and root mean square error (ERMS) less than 0.50 m. The prediction results of the ginkgo tree were the best among the four species (R2 = 0.95, ERMS = 0.31 m). The depth prediction network still had an acceptable performance in the prediction of parameters such as DBH height and crown width, R2 was not less than 0.81, DBH height ERMS was less than 2.50 cm, crown width ERMS was less than 0.32 m. The model in this paper predicted well (R2 ≥ 0.86) with less error compared with other tree parameter prediction methods such as linear regression and LSTM networks.
        Conclusion  The cascaded recurrent neural network can effectively predict the future tree growth and improve the prediction accuracy of tree growth parameters, while the introduction of GRU and attention model has certain robustness in the time-series prediction of growth parameters, which provides a new idea for the intelligent management and visualization analysis of forests.

       

    /

    返回文章
    返回