基于机器学习的青藏高原高寒沼泽湿地蒸散发插补研究

Evapotranspiration interpolation in alpine marshes wetland on the Qingzang Plateau based on machine learning

以青藏高原典型高寒沼泽湿地为观测研究站,以实际蒸散发为研究对象,结合气象因子(净辐射、气温、土壤热通量、风速、相对湿度、土壤含水率),建立基于多元线性回归(MLR)、决策树(CART)、随机森林(RF)、支持向量回归(SVR)、多层感知机(MLP)7种组合5类算法的预测模型,找出对于蒸散发具有较高精度的插补方法,实现实际蒸散发数据集的构建。结果表明:1)研究区蒸散发与净辐射相关性最大,而土壤热通量是影响蒸散发过程的关键因子;2)7种组合的5类机器学习算法模型的决定系数变化范围为0.58-0.83,均方根误差变化范围为0.038-0.089 mm·30 min^(-1);2)随机森林回归模型决定系数最高,模型稳定性最佳,插补效果最优;3)插补完整的蒸散发与净辐射、土壤热通量、气温日尺度变化趋势相同,与风速、相对湿度变化趋势相反。日蒸散发主要集中在生长季,日最大值为8.77 mm·d^(-1),出现在7月9日,日最小值为0.21 mm·d^(-1),出现在1月30日。

Aims This study aims to explore a high-precision interpolation method of evapotranspiration based on machine learning to construct high-quality data set of actual evapotranspiration.Methods Taking the typical alpine marsh wetland on the Qingzang Plateau as the observation station to study evapotranspiration,combined with meteorological factors(net radiation,air temperature,soil heat flux,wind speed,relative humidity,soil volumetric water content),we established a prediction model to construct an actual evapotranspiration data set with a high-precision interpolation method based on combining five methods including multiple linear regression(MLR),decision tree(CART),random forest(RF),support vector regression(SVR)and multi-layer perceptron(MLP).Important findings 1)The correlation between evapotranspiration and net radiation was the largest in the study area,and soil heat flux was the key factor affecting the evapotranspiration process.2)The determination coefficients are from 0.58 to 0.83 among five machine learning algorithm models with seven combinations,and the root mean square error ranges from 0.038 to 0.089 mm·30 min^(-1).3)The random forest regression model has the highest determination coefficient,the best model stability and the best interpolation.4)Interpolated evapotranspiration data had the same diurnal variation trend with net radiation,soil heat flux and ari temperature,but the opposite diurnal variation trend with wind speed and relative humidity.Daily evapotranspiration is mainly concentrated in the growing season,with the daily maximum(8.77 mm)on July 9 and the daily minimum(0.21 mm)on January 30.