基于变量优选与机器学习的农田CO2排放通量反演模型研究

Farmland CO2 Emission Flux Inversion Model Based on Variable Optimization and Machine Learning Algorithm

摘要:

为准确获取农田CO2排放通量和精准监测温室气体排放，采集CO2实测数据，基于光谱影像数据，提取各采样点光谱反射率，在此基础上引入红边波段改进光谱指数，利用变量投影重要性分析（Variable importance in projection，VIP）、皮尔逊相关系数法（Pearson correlation coefficient，PCC）以及灰色关联度分析法（Grey relational analysis，GRA）优选出的特征变量作为模型输入组，基于轻量级梯度提升机（Light gradient boosting machine， LightGBM）、反向传播神经网络（Back-propagation neural network, BPNN）和随机森林（Random forest, RF）机器学习算法，构建36个番茄农田不同生育期CO2排放通量反演模型。结果表明：PCC-GRA变量优选方法构建的模型精度优于VIP和PCC法构建的模型，LightGBM的反演效果整体优于BPNN和RF模型，反演结果能真实反映番茄农田不同生育期CO2排放通量。对比各生育期不同模型反演精度，LightGBM在生长期、开花坐果期、成熟期的反演效果优于其他模型，验证集决定系数R2p分别为0.741、0.818、0.779，均方根误差RMSEp分别为0.035、0.040、0.229mg/(m2·h)，平均绝对误差MAEp分别为0.028、0.034、0.022mg/(m2·h)，其中开花坐果期反演精度表现最优。在果实膨大期，RF反演效果优于其他模型，R2p为0.767，RMSEp为0.031mg/(m2·h)，MAEp为0.360mg/(m2·h)，且基于最佳反演模型PCC-GRA-LightGBM得到的全生育期CO2排放通量动态变化曲线可较为真实地反映研究区CO2排放通量变化特征。研究结果可为农田CO2排放通量的精细化监测与估算提供理论依据。

Abstract:

In order to accurately obtain farmland CO2 emission flux and accurately monitor greenhouse gases, the measured data of CO2 were collected. Based on the spectral image data, the spectral reflectance of each sampling point was extracted, and the red edge band was introduced to improve the spectral index. The feature variables selected by variable importance in projection (VIP), Pearson correlation coefficient (PCC) and grey relational analysis (GRA) were used as the model input group. Based on the lightweight gradient boosting machine (LightGBM), back-propagation neural network (BPNN) and random forest (RF) machine learning algorithms, totally 36 CO2 emission flux inversion models of tomato farmland at different growth stages were constructed. The results showed that the accuracy of the model constructed by PCC-GRA variable selection method was better than that of VIP and PCC methods. The inversion effect of LightGBM was better than that of BPNN and RF models. The inversion results can truly reflect the CO2 emission flux of tomato farmland at different growth stages. Comparing the inversion accuracy of different models in each growth period, the inversion effect of LightGBM in growth period, flowering and fruit setting period and mature period was better than that of other models. The validation set determination coefficients R2p were 0.741, 0.818 and 0.779, respectively, and the root mean square errors (RMSEp) were 0.035mg/(m2·h), 0.040mg/(m2·h) and 0.229mg/(m2·h), respectively. The mean absolute errors (MAEp) were 0.028mg/(m2·h), 0.034mg/(m2·h) and 0.022mg/(m2·h), respectively. The inversion accuracy of the flowering and fruit setting period was the best. In the fruit enlargement period, the RF inversion effect was better than that of other models, R2p was 0.767, RMSEp was 0.031mg/(m2·h), MAEp was 0.360mg/(m2·h), and the dynamic change map of CO2 emission flux in the whole growth period based on the best inversion model PCC-GRA-LightGBM can truly reflect the change characteristics of CO2 emission flux in the study area. The results can provide a theoretical basis for the fine monitoring and estimation of farmland CO2 emission flux.