Please wait a minute...
img

官方微信

遥感技术与应用  2022, Vol. 37 Issue (5): 1190-1197    DOI: 10.11873/j.issn.1004-0323.2022.5.1190
面向双碳的观测与模拟专栏     
基于机器学习和大数据平台的陆地生态系统碳收支遥感监测
高帅1(),侯学会2,汪云3,王倩4,陈悦1,邢瑞1,王晶1,5
1.中国科学院 空天信息创新研究院 遥感科学国家重点实验室,北京 100101
2.山东省农业科学院 农业信息与经济研究所,山东 济南 250100
3.北京林业大学 园林学院,北京 100083
4.天津师范大学 地理与环境科学学院,天津 300387
5.中国地质大学地球科学与资源学院,北京 100083
Remote Sensing Monitoring of Terrestrial Ecosystem Carbon Budget based on Machine Learning and Big Data Platform
Shuai Gao1(),Xuehui Hou2,Yun Wang3,Qian Wang4,Yue Chen1,Rui Xing1,Jing Wang1,5
1.State Key Laboratory of Remote Sensing Science,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101,China
2.Institute of Agricultural Information and Economy,Shandong Academy of Agricultural Sciences,Jinan 250100,China
3.The College of Forestry of Beijing Forestry University,Beijing 100083,China
4.School of Geographic and Environmental Sciences,Tianjin Normal University,Tianjin 300387,China
5.School of Earth Sciences and Resources,China University of Geosciences,Beijing 100083,China
 全文: PDF(2265 KB)   HTML
摘要:

陆地生态系统碳收支是全球碳循环研究的重要指标,也是气候变化的重要参数。针对该指标估测的不确定性,基于陆地生态系统通量观测研究网络的实测碳通量数据及遥感卫星观测数据产品,利用机器学习方法进行建模研究。研究选用随机森林算法自动从高质量的星—地训练数据集中学习特征、挖掘数据中的隐含信息以及时序间依赖关系的差异,建立了基于随机森林算法的碳收支参数GPP(Gross Primary Production)、NEP (Net Ecosystem Production)估算模型,并选择标准指标利用验证数据集对模型进行了客观评价。结果分析表明:与MODIS GPP产品相比,该方法在估算精度上有了提高,其中落叶阔叶林预测结果最优,决策系数为R2为0.82,均方根误差为1.93 gCm-2 d-1,在其他植被类型上也明显优于传统光能利用率模型产品,更接近于地面通量观测数据。基于相同方法建立的NEP模型也得到了较好的估测结果,落叶阔叶林预测模型的输出结果与通量塔获得的NEP相关关系R2为0.70,RMSE=1.75 g C m-2 d-1。GPP和NEP模型精度差异也表明,在进行机器学习建模时,训练数据集自变量的选择仍然需要机理模型支持。为进行陆地生态系统碳收支大范围快速估算,本研究进行了陆地生态系统碳收支遥感监测平台的搭建,该平台以GEE (Google Earth Engine)大数据平台作为数据存储与计算后端,Django和Nginx作为Web服务框架,OpenLayers和jQuery作为前端框架,从而实现了碳收支参数长时间序列大范围的快速计算、结果实时显示等功能。基于该平台和模型获取的2002—2016年全球(60°N—60°S)逐年GPP结果表明,全球平均GPP存在明显的空间差异,显著增加的区域主要集中在亚洲东部地区及北美洲森林地区等。研究表明,基于机器学习和大数据平台进行碳收支参数遥感监测,能够快速提供与地面真实观测较为一致的陆地生态系统区域和全球尺度碳收支遥感监测结果,该流程在一定程度避免了生理过程模型复杂的参数设置,减少了区域和全球大尺度碳收支监测的不确定性。

关键词: 机器学习大数据平台碳收支随机森林时空扩展    
Abstract:

The carbon budget of terrestrial ecosystems is an important indicator of global carbon cycle research and an important parameter of climate change. Based on the terrestrial ecosystem flux observation and remote sensing satellite observation data, machine learning methods are applied for carbon budget estimation. In this study, random forest algorithm is established to automatically learn features from training data and differences in time series dependencies, and carbon related parameters (Gross Primary Production, GPP; Net Ecosystem Production, NEP) could be estimated. Finally, standard indicators are selected to objectively evaluate the model using the validation data set. The result analysis shows that compared with MODIS GPP products, this method has greatly improved the estimation accuracy. Among them, the prediction result of deciduous broad-leaved forest is the best, the decision coefficient R2 is 0.82, and the root mean square error is 1.93 gCm-2 d-1.It is also significantly better than traditional light energy utilization model products in other vegetation types. The NEP machine learning model established based on the same method has also obtained good estimation results. The correlation between the output results of the deciduous broad-leaved forest model prediction model and the NEP obtained by the flux tower is 0.70 and RMSE=1.75 g C m-2 d-1. The difference in accuracy between GPP and NEP models indicates that when machine learning modeling is performed, the selection of independent variables in the training data set still needs to consider theoretical model. In order to quickly estimate the carbon budget of the terrestrial ecosystem, a remote sensing monitoring platform is established. The platform uses the GEE (Google Earth Engine) big data platform as the data storage and computing backend, and Django, HTML, CSS, JavaScript, etc. as the front-end, in order to quick calculation, real-time visualization and other functions. Based on the platform and algorithm, the global (60° N—60° S) GPP results obtained from 2002 to 2016 show that there are obvious spatial differences in the global average GPP, and the significant increase is mainly concentrated in eastern Asia and forested areas in North America. Research shows that remote sensing monitoring of carbon budget parameters based on machine learning and big data platforms can quickly provide regional and global-scale carbon storage and the results are consistent with true ground observations. The obtained estimation results avoid the complicated parameter setting of the physiological process model, and reduce the uncertainty of regional and global large-scale carbon budget monitor.

Key words: Machine learning    Big data platform    Carbon budget    Random forest    Spatio-temporal expansion
收稿日期: 2021-08-17 出版日期: 2022-12-13
ZTFLH:  X16  
基金资助: 国家重点研发计划项目(2017YFA0603004);国家自然科学基金项目(42171377);高分项目(30?Y20A15?9003?17/18);天津市高等学校科技发展计划项目(2018KJ154)
作者简介: 高帅(1983-),男,山东高密人,副研究员,主要从事数据挖掘和激光雷达研究。E?mail:gaoshuai@aircas.ac.cn
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
RSS
作者相关文章  
高帅
侯学会
汪云
王倩
陈悦
邢瑞
王晶

引用本文:

高帅,侯学会,汪云,王倩,陈悦,邢瑞,王晶. 基于机器学习和大数据平台的陆地生态系统碳收支遥感监测[J]. 遥感技术与应用, 2022, 37(5): 1190-1197.

Shuai Gao,Xuehui Hou,Yun Wang,Qian Wang,Yue Chen,Rui Xing,Jing Wang. Remote Sensing Monitoring of Terrestrial Ecosystem Carbon Budget based on Machine Learning and Big Data Platform. Remote Sensing Technology and Application, 2022, 37(5): 1190-1197.

链接本文:

http://www.rsta.ac.cn/CN/10.11873/j.issn.1004-0323.2022.5.1190        http://www.rsta.ac.cn/CN/Y2022/V37/I5/1190

图1  技术路线
图2  模型迭代次数与残差的关系
图3  碳收支监测平台基本架构
图4  云平台Web应用程序整体界面
IGBPGPP_RFRGPP_MODISNEP_RFR
R2

RMSE

(g C m-2 d-1)

R2

RMSE

(g C m-2 d-1)

R2

RMSE

(g C m-2 d-1)

DBF0.812.020.692.690.701.75
GRA0.781.770.62.510.371.44
WSA0.781.120.481.720.411.03
OSH0.710.640.531.170.340.76
CRO0.693.010.414.470.552.37
ENF0.681.920.612.290.351.67
MF0.682.030.612.310.431.64
WET0.612.270.482.690.431.54
EBF0.592.050.442.570.181.90
SAV0.431.870.192.420.241.51
表1  GPP模型、MODIS GPP产品、NEP模型分别与通量塔站点对比
图5  基于机器学习模型的2002—2016年全球平均GPP空间分布
图6  2002—2016年全球GPP
1 Solomon S, Qin D, Manning M, et al. Climate Change 2007:The physical science basis.contribution of working group I to the fourth 16 assessment report of the intergovernmental panel on climate change[R]. Summary for Policy Makers, Geneva,SwitzerLand:Intergovernmental Panel on Climate (IPCC),2007:18.
2 Falkowski P, Scholes R, Boyle E, et al. The global carbon cycle: A test of our knowledge of earth as a system[J]. Science,2000, 290(5490): 291-296.
3 Xia J, Chen J, Piao S, et al. Terrestrial carbon cycle affected by non-uniform climate warming[J]. Nature Geoscience,2014,7(3):173-180.
4 Pourtaghi Z S, Pourghasemi H R, Rossi M. Forest fire susceptibility mapping in the minudasht forests, Golestan Province,Iran[J]. Environmental Earth Sciences,2015,73(4):1515-1533.
5 Running S W, Nemani R R, Heinsch F A, et al. A continuous satellite-derived measure of global terrestrial primary production[J]. Bioscience, 2004, 54(6): 547-560.
6 Raich J W, Schlesinger W H. The Global Carbon-Dioxide Flux in soil respiration and its relationship to vegetation and climate[J].Tellus Series B-Chemical and Physical Meteorology,1992,44(2):81-99.
7 Li Kerang, Huang Mei, Tao Bo.Process-based modeling on the response and adaptation of chinese terrestrial ecosystems to global change[M]. Beijing: China Meteorological Press,2009.
7 李克让, 黄玫, 陶波.中国陆地生态系统过程及对全球变化响应与适应的模拟研究[M].北京:气象出版社,2009.
8 Baldocchi D, Falge E, Gu L H, et al. FLUXNET: A new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and energy flux densities[J]. Bulletin of the American Meteorological Society, 2001, 82(11): 2415-2434.
9 Baldocchi D. Measuring fluxes of trace gases and energy between ecosystems and the atmosphere-the state and future of the eddy covariance method[J].Global Change Biology,2014, 20(12): 3600-360.
10 Geng Shaobo, Lu Shaowei, Rao Liangyi, et al. Research progress of measurement of land surface carbon budget based on eddy covariance technology[J]. World Forestry Research,2010,23(3): 24-28.
10 耿绍波, 鲁绍伟, 饶良懿,等. 基于涡度相关技术测算地表碳通量研究进展. 世界林业研究, 2010,23(3): 24-28.
11 Jensen R, Herbst M, Friborg T. Direct and indirect controls of the interannual variability in atmospheric CO2 exchange of three contrasting ecosystems in denmark[J]. Agricultural and Forest Meteorology,2017,233:12-31.DOI:10.1016/j.agrfor-met2016.10.023 .
doi: 10.1016/j.agrfor-met2016.10.023
12 Zhu A X, Mackay D S. Effects of spatial detail of soil information on watershed modeling[J]. Journal of Hydrology, 2001, 248(1): 54-77.
13 White M A, Thornton P E, S W Running et al. Parameterization and sensitivity analysis of the BIOME-BGC terrestrial ecosystem model: Net primary production controls[J]. Earth Interactions, 2000,4(3):1-84.
14 Chen W, Chen J, Liu J,et al. Approaches for reducing uncertainties in regional forest carbon balance[J]. Global Biogeochemical Cycles, 2000, 14(3): 827-838.
15 Sitch S, Smith B, Prentice I C, et al. Evaluation of ecosystem dynamics, plant geography and terrestrial carbon cycling in the LPJ dynamic global vegetation model[J].Global Change Biology, 2003, 9(2): 161-185.
16 Sims D A, Rahman A F, Cordova V D, et al. A new model of gross primary productivity for North American ecosystems based solely on the Enhanced Vegetation Index and Land Surface Temperature from MODIS[J]. Remote Sensing of Environment, 2008, 112(4): 1633-1646.
17 Gitelson A A, Vina A, Verma S B, et al. Relationship between gross primary production and chlorophyll content in crops: Implications for the synoptic monitoring of vegetation productivity[J]. Journal of Geophysical Research-Atmosphe-res,2006,111:D08S11. DOI:10.1029/2005JD006017 .
doi: 10.1029/2005JD006017
18 Xiao X, Hollinger D, Aber J, et al. Satellite-based modeling of gross primary production in an evergreen needleleaf forest[J]. Remote Sensing of Environment,2004,89(4):519-534.
19 Heinsch F A, Zhao M, Running S W,et al.Evaluation of remote sensing based terrestrial productivity from MODIS using regional tower eddy flux network observations[J].IEEE Tran-sactions on Geoscience and Remote Sensing,2006,44(7): 1908-1925.
20 Yang F, Ichii K, White M A, et al. Developing a continental-scale measure of gross primary production by combining MODIS and ameriflux data through support vector machine approach[J]. Remote Sensing of Environment,2007,110(1):109-122.
21 Jung M, Reichstein M, Bondeau A. Towards global empirical upscaling of fluxnet eddy covariance observations: Validation of a model tree ensemble approach using a biosphere model[J]. Biogeosciences,2009,6(60):2001-2013.
22 Liu X, Guanter L, Liu L, et al. Downscaling of solar-induced chlorophyll fluorescence from canopy level to photosystem level using a random forest model[J]. Remote Sensing of Environment,2019,231:110772.DOI:10.1016/j.rse.2018. 05.035 .
doi: 10.1016/j.rse.2018. 05.035
23 Xu Ziming, Tian Yangfeng. The development history and application of cloud computing[J]. Information Recording Materials, 2018,19(8):66-67.
23 许子明,田杨锋.云计算的发展历史及其应用[J].信息记录材料,2018,19(8):66-67.
24 Chen B, Xiao X, Li X, et al. A mangrove forest map of China in 2015: Analysis of time series Landsat 7/8 and Sentinel-1A imagery in Google Earth Engine cloud computing platform[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 131: 104-120. DOI:10.1016/j.isprsjprs.2017.07.011 .
doi: 10.1016/j.isprsjprs.2017.07.011
25 Alonso A, Munoz-Carpena R, Kennedy R E,et al. Wetland lndscape spatio-temporal degradation dynamics using the new Google Earth Engine cloud-based platform: Opportunities for non-specialists in remote sensing[J]. Transactions of the Asabe, 2016, 59(5): 1331-1342.
26 Chen Y, Shen W, Gao S, et al. Estimating deciduous broadleaf forest gross primary productivity by remote sensing data using a random forest regression model[J]. Journal of Applied Remote Sensing,2019,13(3):038502.DOI:10.1117/1.JRS. 13.038502 .
doi: 10.1117/1.JRS. 13.038502
27 Baldocchi D D.Assessing the eddy covariance technique for eva-luating carbon dioxide exchange rates of ecosystems:Past,pre-sent and future[J].Global Change Biology,2003,9(4):479-492.
28 Papale D, Reichstein M, Aubinet M, et al. Towards a standardized processing of net ecosystem exchange measured with eddy covariance technique: Algorithms and uncertainty estimation[J]. Biogeosciences, 2006, 3(15): 571-583.
29 Hamill T M, Dentremont R P, Bunting J T. A description of the air force real-time nephanalysis model[J]. Weather and Forecasting, 1992, 7(2): 288-306.
30 Hsu K L, Gupta H V, Gao X G,et al.Estimation of physical variables from multichannel remotely sensed imagery using a neural network: Application to rainfall estimation[J]. Water Resources Research, 1999, 35(5): 1605-1618.
31 Zhang K, Liu N, Chen Y, et al. Comparison of different machine learning method for GPP estimation using remote sensing data[J] IOP Conference Series: Materials Science and Engineering,2019,490(6):062010. DOI:10.1088/1757-899X/490/6/062010 .
doi: 10.1088/1757-899X/490/6/062010
32 Gieseke Fabian, Igel Christian. Training big random forests with little resources[C]∥2018:1445-54. DOI:10.1145/3219819.3220124 .
doi: 10.1145/3219819.3220124
33 Chi C, Park T, Wang X,et al. China and India lead in greening of the world through land-use management[J].Nature sustaina-bility,2019,2(2):122-129. DOI:10.1038/s41893-019-0220-7 .
doi: 10.1038/s41893-019-0220-7
[1] 马超,黄华国,田昕,刘炳杰,温坤剑,王鹏杰. 基于背包激光雷达的单木材积估测方法研究[J]. 遥感技术与应用, 2022, 37(5): 1071-1083.
[2] 丁书培,李蒙蒙,汪小钦,李琳,吴瑞姣,黄姮. 基于时序遥感数据的福州市耕地非农化特征及驱动因子分析[J]. 遥感技术与应用, 2022, 37(3): 550-563.
[3] 方攀飞,王雷光,徐伟恒,欧光龙,代沁伶,李若楠. 云南香格里拉地区森林优势树种决策融合分类[J]. 遥感技术与应用, 2022, 37(3): 638-650.
[4] 李媛,周小成,陈芸芝,王锋克. 结合Landsat时序遥感干扰检测的福建将乐县森林林龄估算与评价[J]. 遥感技术与应用, 2022, 37(3): 651-662.
[5] 王美雅,徐涵秋. 超大城市土地覆盖与热环境的随机森林回归模型研究[J]. 遥感技术与应用, 2022, 37(2): 379-388.
[6] 康新礼,张文豪,刘原萍,顾行发,余涛,张丽丽,徐桦昆. 基于随机森林的京津冀地区PM2.5遥感反演及变化分析[J]. 遥感技术与应用, 2022, 37(2): 424-435.
[7] 李永康,王新军,马燕飞,陈蓓,闫立男,张冠宏. 基于机器学习算法的AMSR-2地表温度降尺度研究[J]. 遥感技术与应用, 2022, 37(2): 474-487.
[8] 李翔华,黄春林,侯金亮,韩伟孝,冯娅娅,陈彦四,王静. 结合遥感和统计数据的家畜分布网格化方法研究[J]. 遥感技术与应用, 2022, 37(1): 262-271.
[9] 姜玉峰,齐建国,陈博伟,闫敏,黄龙吉,张丽. 基于无人机高光谱影像和机器学习的红树林树种精细分类[J]. 遥感技术与应用, 2021, 36(6): 1416-1424.
[10] 常潇月,荆林海,林沂. 基于机载LiDAR数据和Landsat 8影像的芬兰北部泰加林—苔原过渡带提取[J]. 遥感技术与应用, 2021, 36(6): 1259-1271.
[11] 魏雪馨,刘洋,闵庆文,刘荣高,张清洋,叶晓星,刘蓓蓓. 基于地物光谱季节曲线特征的毛竹林分布提取[J]. 遥感技术与应用, 2021, 36(5): 1178-1188.
[12] 姚杰鹏,杨磊库,陈探,宋春桥. 基于Sentinel-1,2和Landsat 8时序影像的鄱阳湖湿地连续变化监测研究[J]. 遥感技术与应用, 2021, 36(4): 760-776.
[13] 李晓彤,覃先林,刘倩,赵俊鹏,王崇阳. 基于AISA Eagle Ⅱ机载高光谱数据的森林可燃物类型识别方法[J]. 遥感技术与应用, 2021, 36(3): 544-551.
[14] 盛夏,石玉立,丁海勇. 青藏高原GPM降水数据空间降尺度研究[J]. 遥感技术与应用, 2021, 36(3): 571-580.
[15] 陈彦四,黄春林,侯金亮,韩伟孝,冯娅娅,李翔华,王静. 基于多时相Sentinel-2影像的黑河中游玉米种植面积提取研究[J]. 遥感技术与应用, 2021, 36(2): 324-331.