基于机器学习的蒸散量插补方法

Gap filling method for evapotranspiration based on machine learning

以黑河流域湿地、农田、草地、柽柳、胡杨林、混合林生态系统为研究对象,结合气象因子(净辐射、温度、土壤热通量、风速、相对湿度、土壤体积含水率),分别采用多元线性回归(MLR)、决策树(CART)、随机森林(RF)、支持向量回归(SVR)、BP人工神经网络(BPANN)、深度学习(DL)等方法对蒸散量进行插补。结果表明:(a)RF、SVR、BPANN、DL在各个生态系统的蒸散量插补精度均较高(R 2=0.8~0.93,R MSE=21.730~41.731 W/m 2,M AE=12.153~26.129 W/m 2),但SVR在柽柳、混合林生态系统的结果稍差于其他3种方法(R 2降低了0.01~0.02),MLR插补精度最差(R 2=0.6~0.7),CART结果介于之间(R 2=0.78~0.9)。(b)加入土壤体积含水率能一定程度提升模型插补的精度(R 2提高了0.01~0.06)。(c)利用建立的插补模型去插补其他年份的蒸散量,发现其精度有不同程度的下降。综合考虑模型的精度和稳定性,RF、BPANN、DL对于蒸散量的插补具有较高的精度,同时加入土壤体积含水率可以提高模型插补的精度。

In this study,by selecting the wetland,farmland,grassland,tamarix chinensis,populus euphratica forest and mixed forest ecosystem in the Heihe River Basin as studying objects,several machine learning algorithms were chosen to simulate and interpolate the latent heat flux data,considering several meteorological factors including the net radiation,temperature,soil heat flux,wind speed,relative humidity,and volumetric water content of soil.These methods included multiple linear regression(MLR),decision tree(CART),random forests(RF),support vector regression(SVR),BP artificial neural network(BPANN),and deep learning(DL).The results show the following:(a)RF,SVR,BPANN and DL obtain the best results in different ecosystems with R 2=0.8 to 0.93,R MSE=21.730 to 41.731 W/m 2,and M AE=12.153 to 26.129 W/m 2,but the results of SVR in tamarix and mixed forest ecosystems are slightly worse than other three methods with R 2 decreased by 0.01 to 0.02.The results of MLR are the worst with R 2=0.6 to 0.7 and the results of CART are in between with R 2=0.78 to 0.9.(b)By comparing whether soil moisture participating in the gap filling,it indicates that the participation of soil moisture can improve the gap-filling accuracy of models to some extent with R 2 increased by 0.01 to 0.06.(c)At the same time,the established gap-filling model was used to interpolate the evapotranspiration of other years,and it is found that the accuracy of gap-filling result decreased in varying degrees.With comprehensive consideration of the accuracy and stability of these models,it can be found that RF,BPANN and DL are more suitable for the gap filling of evapotranspiration,and the participation of soil moisture can improve the accuracy of three models.