Modeling Terrestrial Net Ecosystem Exchange Based on Deep Learning in China DOI Creative Commons
Zeqiang Chen, Lei Wu, Nengcheng Chen

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 17(1), P. 92 - 92

Published: Dec. 30, 2024

In estimating the global carbon cycle, net ecosystem exchange (NEE) is crucial. The understanding of mechanism interaction between NEE and various environmental factors ecosystems has been very limited, interactions are intricate complex, which leads to difficulties in accurately NEE. this study, we propose A-DMLP (attention-deep multilayer perceptron)-deep learning model for simulation as well an interpretability study using SHapley Additive exPlanations (SHAP) model. attention was introduced into deep perceptual machine, important information original input data extracted mechanism. Good results were obtained on nine eddy covariance sites China. also compared with random forest, long short-term memory, neural network, convolutional networks (1D) models distinguish it from previous shallow machine estimate NEE, show that have great potential modeling. SHAP method used investigate relationship features simulated enhance normalized difference vegetation index, enhanced leaf area index play a dominant role at most sites. This provides new ideas methods analyzing by introducing interpretable These advancements crucial achieving reduction targets.

Language: Английский

Multiple-Win Effects and Beneficial Implications from Analyzing Long-Term Variations of Carbon Exchange in a Subtropical Coniferous Plantation in China DOI Creative Commons
J. M. Bai,

Fengting Yang,

Huimin Wang

et al.

Atmosphere, Journal Year: 2024, Volume and Issue: 15(10), P. 1218 - 1218

Published: Oct. 12, 2024

To improve our understanding of the carbon balance, it is significant to study long-term variations all components exchange and their driving factors. Gross primary production (GPP), respiration (Re), net ecosystem productivity (NEP) from hourly annual sums in a subtropical coniferous forest China during 2003–2017 were calculated using empirical models developed previously terms PAR (photosynthetically active radiation), meteorological parameters, GPP, Re, NEP calculated. The reasonable agreement with observations, seasonal interannual well reproduced. model-estimated GPP Re over larger than observations 11.38% 5.52%, respectively, model-simulated was lower by 34.99%. showed clear variations, both observed GPPs increased on average 1.04% 0.93%, while values 4.57% 1.06% between 2003 2017. NEPs/NEEs (net exchange) decreased/increased 1.04%/0.93%, which exhibited an increase sink at experimental site. During period 2003–2017, averages air temperature decreased 0.28% 0.02%, water vapor pressure 0.87%. contributed increases NEE 2003–2017. Good linear non-linear relationships found monthly satellite solar-induced fluorescence (SIF) then applied compute relative biases 5.20% 4.88%, respectively. Large amounts CO2 produced clean atmosphere, indicating atmospheric environment will enhance storage plants, i.e., atmosphere beneficial human health sink, as slowing down climate warming.

Language: Английский

Citations

0
Modeling Terrestrial Net Ecosystem Exchange Based on Deep Learning in China DOI Creative Commons
Zeqiang Chen, Lei Wu, Nengcheng Chen

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 17(1), P. 92 - 92

Published: Dec. 30, 2024

In estimating the global carbon cycle, net ecosystem exchange (NEE) is crucial. The understanding of mechanism interaction between NEE and various environmental factors ecosystems has been very limited, interactions are intricate complex, which leads to difficulties in accurately NEE. this study, we propose A-DMLP (attention-deep multilayer perceptron)-deep learning model for simulation as well an interpretability study using SHapley Additive exPlanations (SHAP) model. attention was introduced into deep perceptual machine, important information original input data extracted mechanism. Good results were obtained on nine eddy covariance sites China. also compared with random forest, long short-term memory, neural network, convolutional networks (1D) models distinguish it from previous shallow machine estimate NEE, show that have great potential modeling. SHAP method used investigate relationship features simulated enhance normalized difference vegetation index, enhanced leaf area index play a dominant role at most sites. This provides new ideas methods analyzing by introducing interpretable These advancements crucial achieving reduction targets.

Language: Английский

Citations

0