结合纹理因子和地形因子的森林蓄积量多光谱估测模型

发布时间：2018-04-28 05:19

本文选题：随机森林 + 遥感反演　；参考：《光谱学与光谱分析》2017年07期

【摘要】：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Landsat8OLI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。
[Abstract]:Forest storage is an important factor in the forest survey. It is an important index to evaluate the quantity and quality of forest. The method of measuring the traditional forest accumulation is time-consuming, inefficient, and the precision of the multi linear regression remote sensing inversion method is low. It is difficult to meet the requirements of precision forestry. Machine learning is a kind of use of training data for self improvement and automatic extraction. The method of lifting performance can approximate the nonlinear system arbitrarily and improve the prediction precision of the model. Taking the Jiufeng Forest as the research object, the image spectral factor, the texture factor and the terrain factor are taken into consideration, the BP neural network in machine learning, the least squares support vector machine and the random forest method are used to construct the forest accumulation multispectral estimation model. Type BP-FSV, LSSVM-FSV and RF-FSV, and programmed in Matlab2014a. The aim is to establish two aspects of modeling factor selection and model method, optimize the feature extraction of modeling factor and improve the prediction accuracy of forest storage volume model. The model combined with the Landsat8OLI multi spectral data partition model to predict the forest storage volume inversion modeling. The training ability and prediction ability of the three inversion models are analyzed with the decision coefficient R2 and the mean square root error RMSE as the index. The results show that the combined spectral factors, terrain factors and texture factors are constructed by using the 3 machine learning methods. The sub inversion model can improve the prediction precision of forest storage. In the above model, the RF-FSV model shows strong predictive ability in the three kinds of forest types, which are higher than BP-FSV model, higher or closer to the LSSVM-FSV model.RF-FSV model in the training stage, R2 and RMSE needle Ye Linzhong are 0.839 and 13.9533, and the broad-leaved forest is 0.924 and 7.634. 1, the mixed forest is 0.902 and 12.1539, the prediction stage R2 and RMSE are 0.816 and 15.6301 in the coniferous forest, the broad-leaved forest is 0.913 and 4.8902, the modeling precision and the prediction precision of the mixed forest for 0.865 and 9.344 1.RF-FSV models are high, which provides a new method for forest storage volume remote sensing inversion estimation.

【作者单位】：北京林业大学精准林业北京市重点实验室;平顶山学院旅游与规划学院;中国矿业大学(北京)地球科学与测绘工程学院;
【基金】：国家自然科学基金项目(41371001) 北京市科技专项项目(Z151100001615096) 北京林业大学青年教师科学研究中长期项目(2015ZCQ-LX-01)资助
【分类号】：S758.51;TP79

【相似文献】