基于机器学习与SHAP分析的中国寒温带优势树种分类

Dominant Tree Species Classification in Cold-temperate Forests of China Based on Machine Learning and SHAP Analysis

摘要:

获取较高精度优势树种分布数据是解析北方森林响应气候变化机制的关键。 然而,相比于北美,欧亚北方森林(特别是受气候变化与人为活动影响显著的南缘地带)的树种分布数据仍显匮乏。 为此,本研究选取位于欧亚北方森林南缘的中国寒温带森林为研究区,整合光谱波段、纹理特征、地形因子及生物气候变量等多源数据,应用 XGBoost、Random Forest、MLP 和 Conv1D 4 种机器学习模型,对白桦(Betula platyphylla)、落叶松(Larix gmelinii)和山杨(Populus davidiana)3 种优势树种进行精细识别与绘制。 同时,引入 SHAP 分析从全局和局部视角阐释模型决策机制及特征交互效应。 结果表明:XGBoost 模型性能最优,总体分类精度可达 80. 1% ,分类错误率最低,且山杨在 3 个树种中可分性最佳; SHAP 可解释性分析发现,降水(Bio14)及与温度变化(Bio7、Bio4、Bio3)相关的变量在模型判别中贡献最显著;山杨倾向于温度季节性(Bio4)、日较差(Bio2)较大的环境,在极端气候下具有出色的适应能力,体现了非线性、物种特异性的气候适应性。 研究证实,耦合机器学习、环境因子与模型可解释性分析,可有效解决中国寒温带森林优势树种分类精度不足的问题,增强对气候树种关系的生态解释能力,为制定针对性森林管理策略提供现实依据。

Abstract:

Accurate mapping of dominant tree species distributions is crucial for understanding the response of northern forests to climate change. Compared with North America, species distribution data in Eurasian northern forests, particularly in the southern margin influenced by climate change and human activities, remain scarce. Cold-temperate forests at the southern margin of the Eurasian boreal zone in China were selected as the study area. Multi-source data, including spectral bands, texture features, topographic factors, and bioclimatic variables, were integrated, and four machine learning models, XGBoost, random forest, MLP, and Conv1D were applied to accurately map three dominant species: Betula platyphylla, Larix gmelinii, and Populus davidiana. SHAP analysis was used to interpret model decision mechanisms and feature interactions at both global and local levels. The results indicated that the XGBoost model achieved the highest overall accuracy (80. 1% )with the lowest classification error, and Populus davidiana was the most distinguishable among the three species; SHAP interpretability analysis revealed that precipitation (Bio14)and temperature-related variables (Bio7, Bio4, Bio3 )contributed most to the model discrimination; Populus davidiana occurred preferentially in environments with high temperature seasonality (Bio4 )and diurnal range (Bio2 ), exhibiting strong adaptation under extreme climates, reflecting nonlinear, species-specific responses to climate. The research results demonstrated that integrating machine learning, environmental factors, and model interpretability analysis could effectively address the limited accuracy in classifying dominant tree species in China??s cold-temperate forests, enhance ecological understanding of climate-species relationships, and provide practical guidance for developing targeted forest management strategies.