基于多层感知机模型的长三角水稻种植区净生态系统碳通量模拟  

Simulation of Net Ecosystem Carbon Flux in Rice Planting Area of Yangtze River Delta based on Multi-layer Perceptron Model

作　　者：席闻阳 何建军[1,2] 王智麟 郭立峰 李亚荣[1,4] XI Wenyang;HE Jianjun;WANG Zhilin;GUO Lifeng;LI Yarong(State Key Laboratory of Severe Weather,Chinese Academy of Meteorological Sciences/Key Laboratory of Atmospheric Chemistry,China Meteorological Administration,Beijing 200081,China;College of Marine Science and Engineering,Hainan University,Haikou 570228,Hainan,China;School of Computer Science and Software,Nanjing University of Information Science&Technology,Nanjing 210044,Jiangsu,China;School of Atmospheric Sciences,Lanzhou University,Lanzhou 730000,Gansu,China)

机构地区：[1]中国气象科学研究院,灾害天气国家重点实验室/中国气象局大气化学重点开放实验室,北京200081 [2]海南大学,海洋科学与工程学院,海南海口570228 [3]南京信息工程大学,计算机与软件学院,江苏南京210044 [4]兰州大学大气科学学院,甘肃兰州730000

出　　处：《高原气象》2025年第1期191-200,共10页Plateau Meteorology

基　　金：国家自然科学基金重大项目(42090031);国家自然科学基金面上项目(41975131)。

摘　　要：中国长江三角洲地区(以下简称长三角地区)是典型的水稻种植区,其碳源汇对区域气候和环境具有重要影响。本文系统地分析了长三角地区净生态系统碳交换量(net ecosystem exchange,NEE)与各个气象因子之间的关系,发现NEE与太阳短波辐射的相关性最强,其次与湿度相关参量(饱和水汽压差、相对湿度)等呈现较强的相关性。同时,NEE与太阳辐射、气温、湿度因子、风速和摩擦速度的相关性呈现明显的昼夜变化。基于上述分析,本文利用NEE和气象观测数据构建了长三角水稻下垫面多层感知机(Multilayer perceptron,MLP)NEE模拟模型,评估了模型的模拟效果及其时空稳定性。构建的MLP模型能较好地拟合NEE,训练集模拟的NEE与观测值的相关系数达到0.88,均方根误差为5.34μmol·m^(-2)·s^(-1);MLP模型在模拟长三角水稻季NEE时表现良好,在东台和寿县站点的模拟NEE结果与观测值的相关系数均高于0.78,模型具有较好的时空稳定性;MLP模型模拟白天平均NEE的效果好于夜间平均NEE的效果。研究结果揭示了影响水稻碳循环的主要气象因子,为认识长三角水稻种植区碳循环时空分布特征提供支撑,对准确评估全球和区域碳通量具有重要意义。The Yangtze River Delta in China is a typical rice planting area,and its carbon source and sink have significant impacts on regional climate and environment.This study systematically examines the relationship between NEE and various meteorological factors in the Yangtze River Delta region and reveals that NEE exhibits the strongest correlation with solar short-wave radiation(R=-0.68),followed by a robust linear association with humidity-related parameters(saturated water vapor pressure difference,relative humidity).Additionally,diurnal variations are evident in the correlations between NEE and solar radiation,temperature,humidity factor,wind speed,and friction velocity.Based on these analyses,this paper constructed a multi-layer perceptron(MLP)model for simulating rice undersurface NEE in the Yangtze River Delta using observed NEE data alongside meteorological observations.The simulation performance and spatiotemporal stability of this model are evaluated.Results demonstrate that the constructed MLP model effectively captures NEE patterns;it achieves an R value of 0.88 with respect to observed values within the training set while maintaining an RMSE of 5.34μmol·m^(-2)·s^(-1).Moreover,this MLP model performs well when predicting NEE in the Yangtze River Delta region as evidenced by high correlation coefficients(>0.78)between simulated results and observations at Dongtai and Shouxian stations-indicating good spatiotemporal stability of the model's predictions.Notably,this MLP model demonstrates superior performance when capturing daily variations in daytime mean NEE compared to nighttime mean values.The research results reveal the main meteorological factors affecting rice carbon cycling,provide support for understanding the spatiotemporal distribution characteristics of carbon cycling in rice planting areas of the Yangtze River Delta,and have important significance for accurately evaluating global and regional carbon flux.

关 键 词：机器学习 MLP模型 NEE 长江三角洲地区 水稻种植区 

分 类 号：P401[天文地球—大气物理学与大气环境]