基于清查数据的福建省马尾松生物量转换和扩展因子估算差异解析——3种集成学习决策树模型的比较

生物量转换和扩展因子(BCEFs)是估算森林生物量碳储量普遍使用的重要参数.厘清BCEFs估算差异的来源,可降低森林生物量碳储量评估的不确定性.利用基于集成学习的决策树模型能够很好地解决BCEFs估算差异来源问题.然而,不同此类模型的对比研究目前尚未见报道.本研究以第8次国家森林资源清查福建省331块马尾松的固定样地数据作为材料,分别利用增强回归树(BRT)、随机森林(RF)和立体派(Cubist)模型分析BCEFs(包括地上和地下部分)估算差异的来源.结果表明: 研究区马尾松BCEFs呈右偏分布,平均值为0.69 t·m^-3,最小值为0.67 t·m^-3,最大值为0.71 t·m^-3.BRT、RF和Cubist模型对BCEFs的拟合和预测能力均很好,均能够解释92.8%以上的BCEFs变异.3种模型均给出了相同的前2个相对贡献率最大的自变量,为平均胸径和蓄积量.BCEFs随着平均胸径、蓄积量的增加呈逐渐减小的趋势.平均胸径、蓄积量、平均年龄和平均树高等林分特征因子对BCEFs的影响极大,而土壤因子和地形因子对BCEFs的影响均很小.在建立BCEFs模型时,利用平均胸径、蓄积量、平均年龄和平均树高等少量包含较多BCEFs预测信息的变量便能获取很好的预估精度.当应用固定BCEFs时,应选择在平均年龄、平均胸径以及蓄积等方面具有广泛代表性的样本计算BCEFs.

Difference analysis in estimating biomass conversion and expansion factors of masson pine in Fujian Province,China based on national forest inventory data: A comparison of three decision tree models of ensemble learning.

Biomass conversion and expansion factors (BCEFs) are important parameters for estimating carbon storage in forest biomass. Clarifying the source of differences in estimating BCEFs could reduce uncertainties in forest biomass carbon estimation. The decision tree models of ensemble learning can be used to properly figure out the source of differences in estimating BCEFs. However, the comparison of different decision tree models for analyzing differences in estimating BCEFs has never been reported. In this study, three models the boosted regression trees (BRT), random forest(RF), and Cubist] and data of 331 masson pine plots from the 8^th Chinese National Forest Inventory for Fujian Province were used to analyze the differences in estimating BCEFs (including above- and below-ground). The results showed that BCEFs were following right-skewed distribution, with the mean, minimum and maximum value being 0.69 t·m^-3, 0.67 t·m^-3 and 0.71 t·m^-3, respectively. All three models performed well in BCEFs prediction and fitting, and could explain more than 92.8% variations of BCEFs. All three models showed that average DBH and volume were the top two highest relative importance predictors. BCEFs decreased with the increases of average DBH and volume. Stand characteristics factors, such as average DBH, volume, average age and average height, had great influence on BCEFs. Both soil factors and topographic factors had little influence on BCEFs. Using a few variables (such as average DBH, volume, average age and avera-ge height) which contained more BCEFs prediction information could have preferable forecasting precision when building BCEFs models. Moreover, widely representative samples with different average tree ages, average DBH and volume should be chosen to calculate BCEFs when applying constant BCEFs.