森林蓄积量是林分调查中重要因子, 是评价森林数量和质量的重要指标。 传统森林蓄积量实测方法耗时费力、 效率低下, 多元线性回归遥感反演方法精度较低, 难以达到精准林业要求。 机器学习是一种利用训练数据, 进行自我改进、 自动提升性能的方法, 可以任意逼近非线性系统, 提高模型预测精度。 以鹫峰林场森林为研究对象, 综合考虑影像光谱因子、 纹理因子、 地形因子, 采用机器学习中的BP神经网络、 最小二乘支持向量机、 随机森林方法构建了森林蓄积量多光谱估测模型BP-FSV, LSSVM-FSV和RF-FSV, 并在Matlab2014a中编程实现。 旨在从建模因子选择和模型方法建立两个方面, 优化建模因子特征提取, 提高森林蓄积量模型预测精度。 以角规观测样地实测数据、 森林小班二类调查数据、 林相图数据为基础, 使用以上三种模型结合Landsat8 OLI多光谱数据分林型进行了森林蓄积量反演建模预测。 以决定系数R2和均方根误差RMSE为指标, 分析了三种反演模型的训练能力和预测能力。 研究结果表明: 利用3种机器学习方法构建的结合光谱因子、 地形因子、 纹理因子反演模型能够提高森林蓄积量的预测精度。 以上模型中, RF-FSV模型在针、 阔、 混三种林型中都表现出较强的预测能力, 高于BP-FSV模型, 高于或接近于LSSVM-FSV模型。 RF-FSV模型在训练阶段, R2和RMSE针叶林中为0839和13953 3, 阔叶林中为0924和7634 1, 混交林中为0902和12153 9, 预测阶段R2和RMSE在针叶林中为0816和15630 1, 阔叶林中为0913和4890 2, 混交林中为0865和9344 1。 RF-FSV模型建模精度和预测精度较高, 为森林蓄积量遥感反演估测提供了一种新的方法。

Forest stock volume (FSV) is an important factor in the investigation of the forest stand and the main indicator to evaluate forest. The traditional methods of forest stock volume measurement are time-consuming and low efficiency. In remote sensing multiple linear regression method the accuracy is low and it is difficult to achieve accurate forestry requirements. As a self-improvement and automatic method which using lots of training data, machine learning can approach any nonlinear system model to improve prediction accuracy. Take into account spectral factor, texture factor, topographical factors in study area JIUFENG forest. BP-FSV, LSSVM-FSV and RF-FSV multi-spectral forest volume estimation models were established using BP neural network (BP), least squares support vector machine (LSSVM), random forest (RF) method in machine learning. Ground-angle gauge plots measured data, forest resource in subcompartment inventory data for management, forest sub-compartment map, model in conjunction Landsat8 OLI multispectral remote sensing data of sub-forest types were used for forest volume inversion. Programming in Matlab 2014a realization, BP-FSV Model of BP neural network and LSSVM least squares support vector mechanism LSSVM-FSV model were compared and analyzed based on R2 and RMSE. The results showed that: the p value tested between the predicted values from BP-FSV, LSSVM-FSV and RF-FSV model and observed values is less than 005. It indicates that there is no significant differences between the predicted and observed values of forest stock volume, It shows that the predicted results with the models are ideal, and it is feasible to predict forest stock volume by the models. The model established can improve the forecasting precision of forest stock volume through inversion combining with image spectral, textural, and terrain factor. RF-FSV model in coniferous forest, broad-leaved forest and mixed forest have shown a strong predictive ability, higher than BP -FSV model, which is above or close to LSSVM-FSV model. the RF-FSV model training and predicting accuracy are the highest among the three models, RF-FSV model in the training phase R2 and RMSE is 0839 and 13953 3 in coniferous forest, in broad-leaved forest is 0924 and 7634 1, for mixed forest 0902 and 12153 9. In the prediction stage R2 and RMSE in coniferous forests is 0816 and 15630 1, in broad-leaved forest 0913 and 4890 2, in mixed forest 0865 and 9344 1, it can provide a new method for forest stock volume prediction with better prospects.