基于机器学习的天然混交林单木生长模型

Mixed-species forest has become an important management practice adaptive to climate change, and attained great attention all over the world. Growth modelling for mixed-species stands is still a difficult task in the field of forest growth and yield prediction, however. There is knowledge gap in the growth and inter-specific interaction of mixed-species stands. Climate change is a new challenge to the model. Artificial intelligence development gives a new way to forest growth and yield modelling. Therefore, the project aims to answer the question how to employ machine learning methods to describe the growth and simulate the effects of climate change, forest management and their interaction on the growth of mixed-species forests. Individual tree growth models will be developed based on artificial neural network (ANN) and other 5 assemble machine learning methods (random forest, support vector engine, bagging regression and mboost regression) and traditional statistical models (nonlinear mixed-effects model, generalized additive model). The data come from permanent sample plots，tree ring and dendrometer data. The model will include the effects of competition, climate, stand structure and mixture. Statistical and machine learning based models will be evaluated and compared. The species-specific growth and inter-specific interaction of mixed-species stands will be analyzed from the model. The impact of future climate change and thinning and their interaction will be explored from the simulation with different scenarios. It is expected to provide novel method for mixed-species forest growth modelling and fill the knowledge gap on the growth and adaptive management of mixed conifer-broad-leaved forests.

混交林已经成为应对气候变化的一种重要的适应性森林经营模式，持续得到全球关注。混交林的生长预估一直是森林生长收获预估领域的难点，对混交林的种间关系和生长规律并不完全清楚，并面临气候变化的新挑战。人工智能的发展为混交林生长建模提供了新的思路。本项目探索“如何基于机器学习方法来建立混交林的生长模型，认识混交林的生长规律，并模拟气候变化和森林经营的影响？”这一重要亟待解决的科学问题。以东北温带针阔混交林为对象，将固定样地定期调查数据、树轮数据和直径微生长数据相结合，基于人工神经网络和随机森林等5种机器学习方法，建立包含竞争、气候、结构和混交效应的单木生长模型，并与传统的统计方法进行比较；分析混交林内共存树种间的相互关系和生长规律，模拟间伐和未来气候变化及其交互作用对针阔混交林生长的影响。预期结果将发展混交林生长模拟的新方法，弥补混交林生长规律的知识，为天然针阔混交林的适应性经营提供依据。

混交林已经成为重要的营林体制，其生长预估一直是森林生长收获预估领域的难点，人工智能的发展为混交林生长建模提供了新的思路。本项目以吉林省的针阔混交林为对象，系统开展了基于机器学习的单木和林分生长模型研究，建立了针阔混交林的单木胸径生长、树高-胸径关系、单木冠幅、枯死、单木生物量等模型，在混交林竞竞争指数选择、多输入变量筛选方法、模型相容性和生物合理性等方面取得了创新性成结果。（1）针对影响混交林单木生长的竞争、气候、结构、大小和土壤等因子众多问题，提出了多种变量筛选的综合方法。（2）针对单木竞争指数和竞争木的确定方法问题，通过系统比较发现不同竞争木确定方法与竞争指数有关，但不同方法间的差别很小。（3）针对机器学习算法在森林生长建模应用的过程中输出结果可能会不符合生物学逻辑问题，提出了一个基于理论生长方程的神经网络激活函数，保证了神经网络模型输出的生物合理性。（4）建立了基于深度学习算法的针阔混交林单木冠幅模型，发现了林分结构多样性的正效应及深度神经网络在预测冠幅方面的可靠性。（5）提出了一种基于多任务学习的相容性解决方案，实现了人工神经网络模型相容性无限逼近，可以实现单木生物量各组分与总生物量、单木断面积与林分断面积生长的接近相容性。这些结果证明了机器学习算法在森林生长建模领域的优势和潜力，并在模型相容性方面提出了创新性的解决方案，丰富了森林生长建模理论与方法。项目在Forest Ecology and Management等学科主流期刊上共发表论文6篇，其中SCI收录４篇，培养博士研究生４名，国内外学术交流会上报告４次。培养了森林生长建模领域的人才，推动了机器学习算法在森林生长收获预估领域的应用。