湿地生态系统实际蒸散发数据驱动估算模型研究

The actual evapotranspiration data-driven of wetland ecosystem the estimation model research

利用Fluxnet2015全球通量塔观测数据集,研究了随机森林(RF)、梯度增强回归分析(GBR)、支持向量回归(SVR)和深度学习神经网络(DNN)预测湿地生态系统的实际蒸散发(Evaporation,ET)。通过对比研究,确定了预测实际蒸散发的最佳特征变量组合,包括短波辐射、净辐射、初级生产总值、气温、土壤温度、风速、降水、经度、纬度和时间。以此为模型输入,利用Fluxnet2015站点测试数据集和ERA5-Land再分析资料提供的输入特征,对比分析了不同模型的实际蒸散发估计精度,结果表明:以站点数据为输入,SVR算法精度相对较高,其R^(2)可达0.896,RPE最小为31.5%;以ERA5-Land再分析资料为输入,除了GBR算法以外,其余3种方法R^(2)高于0.820,RPE小于57%。另外,模型算法估计的ET精度要明显高于ERA5-Land再分析资料提供的ET产品。

In this study,random forest(RF),gradient enhanced regression analysis(GBR),support vector regression(SVR)and deep learning neural network(DNN)were used to predict the actual evapotranspiration(Evaporation,ET)of wetland ecosystems using Fluxnet2015 global flux tower observation dataset.Through comparative study,we found that the optimal combination of input features for predicting ET including shortwave radiation,net radiation,gross primary product,air temperature,soil temperature,wind speed,precipitation,longitude,latitude and time.Furthermore,the estimation accuracy of different models was compared and analyzed using independent input datasets extracted from Fluxnet2015 datasets and ERA5-land reanalysis data.The results showed that:taking Fluxnet site data as input,SVR algorithm has a relatively high accuracy,with R^(2)up to 0.896and minimum RPE of 31.5%.Using ERA5-Land reanalysis data as input,except GBR algorithm,the R^(2)of the other three methods was higher than 0.820,RPE was less than 57%.In addition,the accuracies of ET estimated by data-driven algorithms were significantly higher than the ET products in the ERA5-Land reanalysis data.