Chin J Plant Ecol ›› 2023, Vol. 47 ›› Issue (7): 912-921.DOI: 10.17521/cjpe.2022.0015

Special Issue: 生态遥感及应用 生态系统碳水能量通量

• Research Articles • Previous Articles     Next Articles

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

WANG Xiu-Ying, CHEN Qi(), DU Hua-Li, ZHANG Rui, MA Hong-Lu   

  1. Qinghai Institute of Meteorological Science, Key Laboratory of Disaster Prevention and Mitigation of Qinghai Province, Xining 810001, China
  • Received:2022-01-11 Accepted:2022-09-28 Online:2023-07-20 Published:2023-07-21
  • Contact: *CHEN Qi(qq7qq7cq@163.com)
  • Supported by:
    Qinghai Provincial Science and Technology Department Innovation Platform Construction Project(2022-ZJ-Y11);China Meteorological Administration Innovation and Development Project(CXFZ2022P022);China Meteorological Administration Innovation and Development Project(CXFZ2022P046)

Abstract:

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.

Key words: machine learning, alpine marshes, evapotranspiration, cross validation