北京大学学报自然科学版 ›› 2021, Vol. 57 ›› Issue (2): 322-332.DOI: 10.13209/j.0479-8023.2020.121

上一篇    下一篇

基于深度神经网络的城市典型乔木日内蒸腾特征模拟研究

赵文利1, 邱国玉1,†, 熊育久2, 邹振东1, 鄢春华1, 余雷雨1, 郝梦宇1   

  1. 1. 北京大学深圳研究生院环境与能源学院, 深圳 518055 2. 中山大学土木工程学院, 广州 510275
  • 收稿日期:2020-02-08 修回日期:2020-04-10 出版日期:2021-03-20 发布日期:2021-03-20
  • 通讯作者: 邱国玉, E-mail: qiugy(at)pkusz.edu.cn
  • 基金资助:
    深圳市知识创新计划(JCYJ20180504165440088)和国家自然科学基金(41671416)资助

Simulation of Sub-Daily Transpiration Characteristics of Typical Arbor Trees in Cities Based on Deep Neural Network

ZHAO Wenli1, QIU Guoyu1,†, XIONG Yujiu2, ZOU Zhendong1, YAN Chunhua1, YU Leiyu1, HAO Mengyu1   

  1. 1. School of Environment and Energy, Peking University Shenzhen Graduate School, Shenzhen 518055 2. School of Civil Engineering, Sun Yat-Sen University, Guangzhou 510275
  • Received:2020-02-08 Revised:2020-04-10 Online:2021-03-20 Published:2021-03-20
  • Contact: QIU Guoyu, E-mail: qiugy(at)pkusz.edu.cn

摘要:

以城市典型乔木小叶榕全天24小时每10分钟的树干液流及同步气象观测数据为训练集, 建立基于深度神经网络的城市典型乔木植被蒸腾估算模型, 得到10分钟尺度的小叶榕蒸腾模拟结果, 系统地探讨干湿季和昼夜影响小叶榕蒸腾的环境控制因子。基于深圳市91个气象观测站的常规气象观测数据, 应用训练好的深度神经网络模型, 估算得到站点尺度的深圳市典型乔木逐小时蒸腾特征。结果表明: 1) 深度神经网络模型可以高精度地模拟城市小叶榕每10分钟尺度的蒸腾变化, 与树干液流系统实测数据相比, 决定系数R2=0.91, 平均绝对百分比误差MAPE=21.77%, 均方根误差RMSE=0.02 mm/h; 2) 湿季和干季城市小叶榕蒸腾的主要控制因子, 白天均为太阳辐射和气温, 夜间均为饱和水汽压差; 3) 城市小叶榕在夜间仍然存在蒸腾, 干、湿季平均蒸腾速率分别达到0.03和0.01 mm/h; 4) 深圳市不同区域的植被蒸腾特征存在差异, 蒸腾速率最高可相差0.10 mm/h, 总体而言, 湿季白天的蒸腾速率(91个站点均值为0.1 mm/h)比干季白天(均值为0.08 mm/h)更高, 大部分站点夜间植被蒸腾量接近0, 但仍存在蒸腾, 少部分站点干季夜间平均蒸腾速率可达0.07 mm/h, 湿季夜间可达0.10 mm/h。

关键词: 城市蒸散发, 典型乔木, 小叶榕, 深度神经网络, 植被蒸腾, 控制因子

Abstract:

Based on the sap flow system and synchronous meteorological observation data of the typical arbor tree in the city, a transpiration estimation model for urban arbor tree was built using deep neural network. The simulation results can systematically figure out the environmental controlling factors that affect the transpiration of Ficus microcarpa in the dry or wet seasons as well as day or night. Based on the routine meteorological observation data from 91 meteorological observation stations in Shenzhen, the trained deep neural network was used to estimate the station-scale hourly transpiration characteristics of typical arbor trees in Shenzhen. The results show that 1) compared with the measured data of the sap flow system, the deep neural network can accurately simulate the transpiration change of the Ficus microcarpa at 10-minute intervals with a R2 of 0.91, MAPE of 21.77%, RMSE of 0.02 mm/h. 2) The main controlling factors of urban Ficus microcarpa during the wet and dry seasons are solar radiation and air temperature in the daytime, while at night is saturated water vapor pressure deficit. 3) Urban Ficus microcarpa still has transpiration at night, and average value can be 0.03 mm/h and 0.01 mm/h in dry season and wet season, respectively. 4) There are differences among vegetation transpiration in different areas of Shenzhen, with a maximum difference of 0.10 mm/h. In general, the transpiration during the dry season is higher than that during the wet season, and the vegetation transpiration at most sites are close to 0 at night. For some specific sites, the average transpiration at night can reach 0.07 mm/h in dry season, and can reach 0.10 mm/h in the wet season. 

Key words: urban evapotranspiration, typical arbor trees, Ficus microcarpa, deep neural network, vegetation transpiration, control factor