长短期记忆网络在生态系统碳通量反演中的优势与应用

doi:10.12118/j.issn.1000-6060.2025.092

摘要/Abstract

摘要：

基于长短期记忆网络方法，结合FLUXNET通量站观测数据与对应遥感生物物理参数数据，提出通过综合欧氏距离指数量化训练集与测试集之间数据异质性，并结合模糊数学理论构建碳通量反演模型。结果表明：（1）通过数据预处理、模型训练，分别利用随机森林、支持向量机、多元线性回归和长短期记忆网络算法建立模型，发现长短期记忆网络在碳通量反演中具有优势。（2）采用留一交叉验证策略，在不同气候区内构建碳通量反演模型，以反映地表空间异质性，并用决定系数（R²）对模型进行评估，发现综合欧式距离与R²之间为显著负相关。（3）将构建的模型应用于美国通量站进行验证，总初级生产力和生态系统呼吸的R²均值均为0.72。总体而言，研究结果提出了一种有效的碳通量模拟方法，具有较好的应用潜力。

关键词: 机器学习, 长短期记忆网络, 遥感, 碳通量, 模糊数学

Abstract:

This study proposes a carbon flux inversion model based on the long short-term memory (LSTM) network. A comprehensive Euclidean distance index is introduced by integrating FLUXNET flux tower observation data with corresponding remote-sensing biophysical parameter datasets to quantify data heterogeneity between training and testing sets. Furthermore, a fuzzy mathematics theory is incorporated to develop the inversion model. Models were developed using random forest, support vector machine, multiple linear regression, and LSTM algorithms through data preprocessing and model training. Results revealed that the LSTM network performed better than the other algorithms in carbon flux inversion. In addition, using the leave-one cross-validation strategy, many carbon flux machine learning models were developed to reflect the spatial heterogeneity of the surface, and the determination coefficient R² was used to evaluate the models. Results revealed that the comprehensive Euclidean distance was significantly negatively correlated with R². The constructed model was applied to the US flux station for verification, and the mean R² values of the total primary productivity and ecosystem respiration were both 0.72. Overall, this study proposed an effective carbon flux simulation method, which has good application potential.

Key words: machine learning, long short-term memory, remote sensing, carbon flux, fuzzy mathematics

高瑞翔, 罗格平, 张文强, 谢明娟, 王渊刚. 长短期记忆网络在生态系统碳通量反演中的优势与应用[J]. 干旱区地理, 2025, 48(12): 2210-2219.

GAO Ruixiang, LUO Geping, ZHANG Wenqiang, XIE Mingjuan, WANG Yuangang. Ecosystem carbon flux inversion method combining LSTM and fuzzy mathematics[J]. Arid Land Geography, 2025, 48(12): 2210-2219.

图/表 9

表1

表2

图1

图2

图3

表3

图4

图5

图6

参考文献 46

[1]	Kanevski M, Pozdnoukhov A, Timonin V. Machine learning for spatial environmental data: Theory, applications, and software[M]. New York: EPFL Press, 2009: 40-52.
[2]	朴世龙, 何悦, 王旭辉, 等. 中国陆地生态系统碳汇估算: 方法、进展、展望[J]. 中国科学: 地球科学, 2022, 52(6): 1010-1020.
	[Piao Shilong, He Yue, Wang Xuhui, et al. Estimation of carbon sinks in terrestrial ecosystems in China: Methods, progress, and prospects[J]. Scientia Sinica (Terrae), 2022, 52(6): 1010-1020.]
[3]	Zhan W W, Yang X, Ryu Y, et al. Two for one: Partitioning CO₂ fluxes and understanding the relationship between solar-induced chlorophyll fluorescence and gross primary productivity using machine learning[J]. Agricultural and Forest Meteorology, 2022, 321: 108980, doi: 10.1016/j.agrformet.2022.108980.
[4]	Pichler M, Hartig F. Machine learning and deep learning: A review for ecologists[J]. Methods in Ecology and Evolution, 2023, 14(4): 994-1016. doi: 10.1111/mee3.v14.4
[5]	Reichstein M, Camps-Valls G, Stevens B, et al. Deep learning and process understanding for data-driven Earth system science[J]. Nature, 2019, 566(7743): 195-204. doi: 10.1038/s41586-019-0912-1
[6]	Hungershoefer K, Breon F-M, Peylin P, et al. Evaluation of various observing systems for the global monitoring of CO₂ surface fluxes[J]. Atmospheric Chemistry and Physics, 2010, 10(21): 10503-10520.
[7]	Ciais P, Peylin P, Bousquet P. Regional biospheric carbon fluxes as inferred from atmospheric CO₂ measurements[J]. Ecological Applications, 2000, 10(6): 1574-1589.
[8]	Foley J A, Prentice I C, Ramankutty N, et al. An integrated biosphere model of land surface processes, terrestrial carbon balance, and vegetation dynamics[J]. Global Biogeochemical Cycles, 1996, 10(4): 603-628. doi: 10.1029/96GB02692
[9]	Houghton R A, House J I, Pongratz J, et al. Carbon emissions from land use and land-cover change[J]. Biogeosciences, 2012, 9(12): 5125-5142. doi: 10.5194/bg-9-5125-2012
[10]	Liang S. Quantitative remote sensing of land surfaces[M]. Canada: John Wiley & Sons, 2005: 10-22.
[11]	Liu X, Lu D, Zhang A, et al. Data-driven machine learning in environmental pollution: Gains and problems[J]. Environmental Science & Technology, 2022, 56(4): 2124-2133. doi: 10.1021/acs.est.1c06157
[12]	Tuia D, Kellenberger B, Beery S, et al. Perspectives in machine learning for wildlife conservation[J]. Nature Communications, 2022, 13(1): 792, doi: 10.1038/s41467-022-27980-y. pmid: 35140206
[13]	Silva L A, Zanella G. Robust leave-one-out cross-validation for high-dimensional Bayesian models[J]. Journal of the American Statistical Association, 2023, 119(547): 2369-2381. doi: 10.1080/01621459.2023.2257893
[14]	Watson G L. On model determination, prediction and statistical learning: The case of space-time data[M]. California: University of California, Los Angeles, 2021: 1-33.
[15]	Meyer H, Pebesma E. Predicting into unknown space? Estimating the area of applicability of spatial prediction models[J]. Methods in Ecology and Evolution, 2021, 12(9): 1620-1633. doi: 10.1111/mee3.v12.9
[16]	Xu Z, Peng J, Dong J, et al. Spatial correlation between the changes of ecosystem service supply and demand: An ecological zoning approach[J]. Landscape and Urban Planning, 2022, 217: 104258, doi: 10.1016/j.landurbplan.2021.104258.
[17]	Shafahi A, Huang W R, Najibi M, et al. Poison frogs! targeted clean-label poisoning attacks on neural networks[J]. Advances in Neural Information Processing Systems, 2018, 31: 42, doi: 10.48550/arXiv.1804.00792.
[18]	Singla A, Bertino E, Verma D. Preparing network intrusion detection deep learning models with minimal data using adversarial domain adaptation[R]. United States:Proceedings of the 15^th ACM Asia Conference on Computer and Communications Security, 2020.
[19]	Gu X, Easwaran A. Towards safe machine learning for cps: Infer uncertainty from training data[R]. Montreal, QC, Canada:In 10^thACM/IEEE International Conference on Cyber-Physical Systems, 2019.
[20]	Zhu A X, Lu G, Liu J, et al. Spatial prediction based on Third Law of Geography[J]. Annals of GIS, 2018, 24(4): 225-240. doi: 10.1080/19475683.2018.1534890
[21]	Zhu A, Lü G, Zhou C, et al. Geographic similarity: Third law of geography[J]. Journal of Geo-information Science, 2020, 22(4): 673-679.
[22]	Jegou H, Douze M, Schmid C. Product quantization for nearest neighbor search[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2010, 33(1): 117-128. doi: 10.1109/TPAMI.2010.57
[23]	Sheridan R P, Feuston B P, Maiorov V N, et al. Similarity to molecules in the training set is a good discriminator for prediction accuracy in QSAR[J]. Journal of Chemical Information and Computer Sciences, 2004, 44(6): 1912-1928. pmid: 15554660
[24]	于贵瑞, 王秋凤, 朱先进. 区域尺度陆地生态系统碳收支评估方法及其不确定性[J]. 地理科学进展, 2011, 30(1): 103-113. doi: 10.11820/dlkxjz.2011.01.013
	[Yu Guirui, Wang Qiufeng, Zhu Xianjing. Methods and uncertainties in evaluating the carbon budgets of regional terrestrial ecosystems[J]. Progress in Geography, 2011, 30(1): 103-113.] doi: 10.11820/dlkxjz.2011.01.013
[25]	李宏业, 谢惠春. 基于涡度相关法的不同生态系统碳通量研究进展[J]. 青海草业, 2024, 33(3): 33-38.
	[Li Hongye, Xie Huichun. Research progress on carbon fluxes in different ecosystems based on eddy correlation method[J]. Qinghai Prataculture, 2024, 33(3): 33-38.]
[26]	Massman W J, Lee X. Eddy covariance flux corrections and uncertainties in long-term studies of carbon and energy exchanges[J]. Agricultural and Forest Meteorology, 2002, 113(1): 121-144. doi: 10.1016/S0168-1923(02)00105-3
[27]	Yu G, Fu Y, Sun X, et al. Recent progress and future directions of ChinaFLUX[J]. Science in China Series D: Earth Sciences, 2006, 49(Suppl. 2): 1-23.
[28]	陈世苹, 游翠海, 胡中民, 等. 涡度相关技术及其在陆地生态系统通量研究中的应用[J]. 植物生态学报, 2020, 44(4): 291-304. doi: 10.17521/cjpe.2019.0351
	[Chen Shiping, You Cuihai, Hu Zhongmin, et al. Eddy covariance technique and its applications in flux observations of terrestrial ecosystems[J]. Chinese Journal of Plant Ecology, 2020, 44(4): 291-304.] doi: 10.17521/cjpe.2019.0351
[29]	岳斌, 余赛芬, 董晶晶, 等. 温室气体通量测量方法及进展[J]. 光学学报, 2023, 43(18): 90-102.
	[Yu Bing, Yu Saifen, Dong Jingjing, et al. Measurement methods and progress of greenhouse gas flux[J]. Optics Journal, 2023, 43(18): 90-102.]
[30]	Knyazikhin Y. MODIS leaf area index (LAI) and fraction of photosynthetically active radiation absorbed by vegetation (FPAR) product (MOD15) algorithm theoretical basis document[DB/OL]. [1999-04-30]. http://eospso gsfc nasa gov/atbd/modistables html.
[31]	Myneni R B, Hoffman S, Knyazikhin Y, et al. Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data[J]. Remote Sensing of Environment, 2002, 83(1-2): 214-231. doi: 10.1016/S0034-4257(02)00074-3
[32]	Ryu Y, Jiang C, Kobayashi H, et al. MODIS-derived global land products of shortwave radiation and diffuse and total photosynthetically active radiation at 5 km resolution from 2000[J]. Remote Sensing of Environment, 2018, 204: 812-825. doi: 10.1016/j.rse.2017.09.021
[33]	Hastings D, Dunbar P, Elphingstone G, et al. SAFARI 2000 digital elevation model, 1-km (GLOBE)[J]. ORNL Distributed Active Archive Center (DAAC) Dataset, 2002, 630, doi: 10.3334/ORNLDAAC/630.
[34]	Zhang W, Luo G, Chen C, et al. Quantifying the contribution of climate change and human activities to biophysical parameters in an arid region[J]. Ecological Indicators, 2021, 129: 107996, doi: 10.1016/j.ecolind.2021.107996.
[35]	Xiao J F, Zhuang Q L, Law B E, et al. Assessing net ecosystem carbon exchange of US terrestrial ecosystems by integrating eddy covariance flux measurements and satellite observations[J]. Agricultural and Forest Meteorology, 2011, 151(1): 60-69. doi: 10.1016/j.agrformet.2010.09.002
[36]	Gorelick N, Hancher M, Dixon M, et al. Google Earth Engine: Planetary-scale geospatial analysis for everyone[J]. Remote Sensing of Environment, 2017, 202: 18-27. doi: 10.1016/j.rse.2017.06.031
[37]	Kulanuwat L, Chantrapornchai C, Maleewong M, et al. Anomaly detection using a sliding window technique and data imputation with machine learning for hydrological time series[J]. Water, 2021, 13(13): 1862, doi: 10.3390/w13131862.
[38]	Nash D J. World atlas of desertification[J]. The Geographical Journal, 1999, 165: 325-326. doi: 10.2307/3060449
[39]	Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780. doi: 10.1162/neco.1997.9.8.1735 pmid: 9377276
[40]	Graves A, Graves A. Supervised sequence labelling with recurrent neural networks[M]. New York Dordrecht London: Springer-Verlag Berlin Heidelberg, 2012: 37-45.
[41]	Ancona M, Ceolini E, Öztireli C, et al. Explainable AI:Interpreting, explaining and visualizing deep learning[M]. Switzerlan: Springer Nature Switzerland AG, 2019: 169-191.
[42]	Zhu A X. Measuring uncertainty in class assignment for natural resource maps under fuzzy logic[J]. Photogrammetric Engineering and Remote Sensing, 1997, 63(10): 1195-1201.
[43]	Fan N Q, Zhao F H, Zhu L J, et al. Digital soil mapping with adaptive consideration of the applicability of environmental covariates over large areas[J]. International Journal of Applied Earth Observation and Geoinformation, 2022, 113: 102986, doi: 10.1016/j.jag.2022.102986.
[44]	Farahani M A, Goodwell A E. Causal drivers of land-atmosphere carbon fluxes from machine learning models and data[J]. Journal of Geophysical Research: Biogeosciences, 2024, 129(6): e2023JG 007815, doi: 10.1029/2023JG007815.
[45]	Huang C, He W, Liu J, et al. Exploring the potential of long short-term memory networks for predicting net CO₂ exchange across various ecosystems with multi-source data[J]. Journal of Geophysical Research: Atmospheres, 2024, 129(7): e2023JD040418, doi: 10.1029/2023JD040418.
[46]	Novick K A, Biederman J, Desai A, et al. The AmeriFlux network: A coalition of the willing[J]. Agricultural and Forest Meteorology, 2018, 249: 444-456. doi: 10.1016/j.agrformet.2017.10.009

解释变量	描述	单位	时间/空间分辨率	来源
T_max	最高气温	℃	日	通量塔
T_min	最低气温	℃	日	通量塔
T_mean	平均气温	℃	日	通量塔
P	降水量	mm	日	通量塔
WS	风速	m·s^-1	日	通量塔
VPD	大气水气压差	hPa	日	通量塔
LAI	叶面积指数	m²	8 d/500 m	MODIS
DSR	向下短波辐射	W·m^-2	d/5 km	Seoul National
DEM	高程	m	静态变量	GLOBE from NOAA
Aspect	坡向	(°)	静态变量	GLOBE from NOAA
Slope	坡度	(°)	静态变量	GLOBE from NOAA
Clay	表土黏土占比	%	静态变量	HWSD from FAO
Sand	表土砂占比	%	静态变量	HWSD from FAO
Silt	表土粉砂占比	%	静态变量	HWSD from FAO

干旱指数分区	LSTM模型	RF模型	SVM模型	MLR模型
干旱区（GPP）	0.62（±0.37）	0.80（±0.20）	0.98（±0.55）	0.99（±0.32）
干旱区（ER）	0.78（±0.40）	0.99（±0.65）	1.01（±0.69）	1.17（±0.66）
半干旱区（GPP）	0.74（±0.50）	0.81（±0.52）	0.84（±0.33）	0.80（±0.51）
半干旱区（ER）	0.85（±1.13）	0.96（±1.07）	0.87（±0.76）	0.90（±0.91）
半湿润区（GPP）	1.06（±2.04）	1.31（±2.90）	1.29（±2.64）	2.99（±9.24）
半湿润区（ER）	0.80（±1.14）	0.84（±0.94）	0.90（±1.45）	1.95（±1.05）
湿润区（GPP）	0.61（±0.39）	0.63（±0.34）	0.78（±0.29）	0.70（±0.55）
湿润区（ER）	0.63（±0.45）	0.64（±0.51）	0.83（±0.46）	0.78（±0.70）