-
植被对于气候变化、人类生存和社会发展都具有重大意义。遥感影像已成为提供植被状况连续信息的重要技术手段[1],传感器的红光和红外波段可以反映高达90%的植被信息[2],常被用于构建遥感植被指数,以实现大范围的植被动态监测。目前,归一化植被指数(NDVI)是应用最多的植被指数之一,被证明在地表植被调查[3-4]、碳循环监测[5-6]、作物产量评估[7]、荒漠化研究[8-9]等方面均有较好的应用。遥感技术发展近半个世纪,卫星传感器仍不得不在时间和空间分辨率之间做出权衡,少有数据能同时兼具高时空分辨率的特征[10]。最为典型的Landsat系列卫星数据,其多光谱波段影像空间分辨率为30 m,被广泛应用于植被覆盖类型制图及状况调查[11-14],但其16 d的重访周期,加之云雨天气影响的延长,严重影响了其在植被动态监测方面的应用[15]。而MODIS数据的植被产品具有很好的一致性及多时相的特点,在植被物候、状况动态等监测中有良好的应用[16-20],其最高250 m的空间分辨率,难以捕捉较小区域内的空间特征差异和满足精细化的植被监测管理[21]。为实现高精度的地表植被状况监测,研究人员提出了多源遥感数据时空融合,即通过融合高空间分辨率和高时间分辨率遥感数据,获得高时空分辨率数据[22]。
不同时空融合方法从不同角度出发,在不同研究区域获得了较好的融合效果,但是各方法之间的差异及其适用性还有待深入研究。石月禅等[23]以盈科灌溉区域为例,利用多时相MODIS数据和高空间分辨率的ASTER/TM影像,比对了基于时序数据的时空数据融合(STIFM)、基于混合像元分解的时空数据融合(STDFM)和基于增强型时空自适应反射率融合(ESTARFM)等3种模型,认为对于NDVI数据的融合,ESTARFM在异质性较强区域具有更好的适用性。HOBYB等[24]比对了时空自适应反射率融合(STARFM)、ESTARFM和灵活的时空数据融合(FSDAF)等3种模型融合生成高时空分辨率NDVI数据的效果,认为ESTARFM相对于另外2种模型融合结果更为准确,同时对于输入数据质量的敏感性较低,具有较高的稳定性。ZHOU等[25]比较了6种典型的时空融合模型,包括基于分解的数据融合(UBDF)、线性混合增长模型(LMGM)、STARFM、回归拟合空间滤波和残差补偿模型(Fit-FC)、一对字典学习模型(OPDL)、灵活时空数据融合模型(FSDAF),并推荐由WANG等[26]提出的Fit-FC模型用于NDVI影像的时空数据融合。然而,这些研究多集中在不同模型的空间细节特征融合效果的比较,而少有关注不同模型的动态特征模拟效果。
三江源区域位于亚欧大陆中纬度地区,是全球气候变化最为敏感的生态区域之一,还是中国重要的生态缓冲区和生态系统服务功能区[27-28],因而该区域的植被状况一直受到研究人员的重点关注[29-31]。本研究在三江源地区选取了2块地表特征具有一定差异的区域,比较STARFM、ESATARM、Fit-FC和规则集回归树融合模型(RPRTM)等4种不同遥感数据融合模型在NDVI时空融合中的应用能力。并以真实的Landsat影像为参考,通过定性的目视判别和定量的统计分析来评价不同融合模型结果的空间特征融合效果,同时将融合结果与MODIS时间序列NDVI进行比较,深入讨论不同融合模型的优点及适用性,及时准确地获取三江源地区生长季内连续的时空高分辨率数据,以便进行地表植被状况监测。
-
图2为以6月28日为目标时刻,STARFM、ESTARFM和Fit-FC和RPRTM生成的NDVI结果,以及对应的MODIS NDVI真值(6月26日)、Landsat NDVI真值(6月28日)。通过目视解译分析,在2个不同区域,这4种融合模型结果均能在一定程度上显示较高分辨率的空间分布特征。从区域1的Landsat NDVI真实影像(图2A)可以看出:该区域具有较为丰富的纹理特征,地块之间边界清晰,与MODIS NDVI影像( 图2B)空间格局基本一致。STARFM融合结果(图2C)中斑块化问题较明显,耕地边界出现锯齿状模糊不清,ESTARFM (图2D)和Fit-FC (图2E)的融合结果要明显优于STARFM,可以清晰看出耕地、草地和沙地等不同地物的空间分布,与Landsat真实影像相似度很高。该区域RPRTM融合结果纹理特征更为清晰(图2F),可以清楚地看出不同耕地区域的边界,与Landsat NDVI真实影像一致性高,融合结果较好。
从区域2的Landsat NDVI真实影像中可以看到明显的地形起伏特征,以及清晰的河流边界(图2G)。MODIS NDVI影像的空间格局与Landsat基本一致,但森林区域的NDVI像元值略微偏低(图2H)。STARFM融合结果与Landsat NDVI影像的地形起伏格局基本一致,但是河流边界出现了若干的斑块问题(图2I)。ESTARFM融合结果具有清晰的地形变化特征,更接近Landsat NDVI真实影像,河流边界清晰可见(图2J)。Fit-FC融合结果(图2K)与Landsat真值影像相似,且与ESTARFM相近。RPTRM融合结果同样地形细节特征清晰,河流边界明显(图2L)。但与ESTARFM不同,RPRTM是以MODIS像元值为目标通过站点训练构建的融合模型,因而其融合结果像元值更接近于MODIS像元值,略低于Landsat影像的NDVI值。
-
从表1和图3可以看出:在区域1中,RPRTM与Landsat NDVI真值的R2最高(0.82),MAD (0.04)和RMSE (0.04)相对较小,表明在该模型下预测图像所含信息丰富,效果最佳。其次为Fit-FC,与Landsat NDVI真值的R2为0.76,MAD、RMSE、Std、AG、IE分别为0.03、0.05、0.09、0.01、6.07;区域2的情况有所不同,ESTARFM与Landsat NDVI真值的R2最高,为0.95,MAD和RMSE最小,均为0.02,表明在该区域ESTARFM的融合结果与Landsat真实影像的相似度最高。造成这种差异的原因主要是相对于区域1,区域2模型输入数据(Landsat和MODIS)差异较大。STARFM、ESTARFM和Fit-FC同属于基于重构的多源遥感数据时空融合方法,根据光谱线性混合原理,通过2期MODIS影像的差异来模拟目标日期的Landsat,其融合结果与Landsat真值更为接近。而RPRTM则属于基于学习的多源遥感数据时空融合模型,以MODIS NDVI为目标变量进行模型训练,融合后结果与MODIS真值更为接近。所以当目标日期的Landsat影像与MODIS值差异较大时,RPRTM融合结果与Landsat影像的相似度相对较低。Fit-FC在区域1和区域2中与Landsat真实影像均有较高相似度,R2分别为0.76、0.90,表明该模型对多种地表覆盖状况的多源遥感数据融合有较好的适用性。
表 1 各模型融合结果比较
Table 1. Comparison of different model result
区域 融合模型 R2 MAD RMSE Std AG IE 区域1 Landsat 0.01 6.31 MODIS 0.02 6.33 STARFM 0.60 0.04 0.07 0.10 0.01 6.17 ESTARFM 0.66 0.04 0.06 0.09 0.01 6.08 Fit-FC 0.76 0.03 0.05 0.09 0.01 6.07 RPRTM 0.82 0.04 0.04 0.10 0.02 6.34 区域2 Landsat 0.02 6.13 MODIS 0.01 5.97 STARFM 0.88 0.02 0.07 0.09 0.01 5.92 ESTARFM 0.95 0.02 0.02 0.10 0.02 6.13 Fit-FC 0.90 0.17 0.18 0.15 0.03 6.92 RPRTM 0.62 0.03 0.06 0.07 0.02 6.02 -
图4包括了MODIS和Landsat的NDVI观测数据以及4种模型融合数据的NDVI时序统计结果,时间分辨率为16 d,从2013年的第129天(5月9日)到第273天(9月30日)。总体来看,对于3种不同植被类型,4种融合方法均能较好地模拟其季节动态特征,能准确反映不同植被的生长动态变化。融合后的NDVI时间序列波动趋势与MODIS NDVI基本一致,在生长季内呈明显的单峰特征。其中RPRTM融合后NDVI曲线与MODIS真值吻合度最高,几乎重叠,其次为Fit-FC、ESTARFM,最后为STARFM,表明RPRTM融合后的NDVI时间序列最接近MODIS真实值。通过相关性分析同样可以看出:对于3种不同地表植被,RPRTM融合结果均取得了与MODIS NDVI最高的相关性,在草地(R2=0.99)、耕地(R2=0.99)和森林(R2=0.97)区域,对于地表植被状况的季节动态捕捉与MODIS真实值保持着高度的一致性。
Comparison of four fusion models for generating high spatio-temporal resolution NDVI
-
摘要:
目的 针对时空融合方法在遥感植被状况调查及动态变化监测中的应用,比对时空自适应反射率融合模型(STARFM)、增强型时空自适应反射率融合模型(ESTARFM)、回归拟合空间滤波和残差补偿模型(Fit-FC)和规则集回归树融合模型(RPRTM)等4种时空融合模型对归一化植被指数(NDVI)的融合效果。 方法 以三江源地区2块具有差异性地表特征的区域为研究样地,采用上述4种时空融合方法,融合空间分辨率30 m的Landsat 8影像和250 m时间步长16 d的MODIS NDVI数据,生成步长为16 d的30 m空间分辨率的NDVI数据。基于Landsat NDVI影像通过定性的目视判别和定量的统计分析来评价不同融合模型结果的空间特征模拟效果,并以真实的MODIS NDVI时间动态为参考,分析了不同融合方法对地表植被动态特征的拟合效果。 结果 ①关于空间特征的捕捉,在地表覆盖状况较复杂的区域,RPRTM融合效果最佳(R2=0.82);而对于输入影像差异较大的区域,ESTARFM融合效果最佳(R2=0.95)。②关于时间动态的捕捉,RPRTM针对不同的植被型均取得了最佳效果(R2为0.97~0.99)。③相对于模型输入数据的时空可比性,地表异质性对STARFM和ESTARFM融合效果的影响更大。 结论 4种时空融合模型能有效用于生成高时空分辨率的NDVI数据,不同模型其融合效果各有不同,RPRTM在复杂地表区域与模拟植被生长动态变化中均有较好表现。图4表1参38 -
关键词:
- 时空数据融合 /
- 归一化植被指数 /
- 增强型时空自适应反射率融合模型 /
- 规则集回归树融合模型 /
- 回归拟合空间滤波和残差补偿模型
Abstract:Objective In order to choose adapted fusion methods in vegetation survey and dynamic monitoring, we applied four different spatio-temporal fusion models including spatial and temporal adaptive reflectance fusion model (STARFM), enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), regression model fitting, spatial filtering and residual compensation (Fit-FC) and the rule-based piecewise regression tree model (RPRTM). Method Based on the four spatio-temporal fusion models (STARFM, ESTARFM, Fit-FC and RPRTM), two sampling regions (region Ⅰ and Ⅱ), with different surfaces characteristics in the Three-River Headwaters Regions were taken to generate the high spatial information of the Landsat NDVI (30 m, 16 d). Based on Landsat NDVI image, the spatial characteristics of the fusion data of different fusion models were evaluated by qualitative visual discrimination and quantitative statistical analysis. Meanwhile, based on the MODIS NDVI time series, the fitting effect of different fusion methods on the dynamic characteristics of surface vegetation was analyzed. Result (1) RPRTM had the optimal spatial fusion performance in region Ⅰ (R2=0.82); and ESTARFM performed the best in region Ⅱ (R2=0.95). (2) RPRTM has achieved the best fusion for capturing temporal dynamics (R2=0.97−0.99), where the NDVI dynamics were highly consistent with the temporal variations of MODIS. (3) Compared with the spatio-temporal comparability of model input data, landscape heterogeneity had a greater impact on the fusion effect of STARFM and ESTARFM. Conclusion Spatio-temporal fusion models can be used effectively to generate NDVI data at high spatial and temporal resolution, with different models having different fusion effects. RPRTM performing well in both complex surface areas and simulated vegetation growth dynamics. [Ch, 4 fig. 1 tab. 38 ref.] -
表 1 各模型融合结果比较
Table 1. Comparison of different model result
区域 融合模型 R2 MAD RMSE Std AG IE 区域1 Landsat 0.01 6.31 MODIS 0.02 6.33 STARFM 0.60 0.04 0.07 0.10 0.01 6.17 ESTARFM 0.66 0.04 0.06 0.09 0.01 6.08 Fit-FC 0.76 0.03 0.05 0.09 0.01 6.07 RPRTM 0.82 0.04 0.04 0.10 0.02 6.34 区域2 Landsat 0.02 6.13 MODIS 0.01 5.97 STARFM 0.88 0.02 0.07 0.09 0.01 5.92 ESTARFM 0.95 0.02 0.02 0.10 0.02 6.13 Fit-FC 0.90 0.17 0.18 0.15 0.03 6.92 RPRTM 0.62 0.03 0.06 0.07 0.02 6.02 -
[1] 杨俊. 基于ETM+遥感影像的森林覆盖面积提取方法的研究[D]. 南京: 南京农业大学, 2006. YANG Jun. Study of Extracting the Forest Area Based on the ETM+ Remote Sensing Image [D]. Nanjing: Nanjing Agricultural University, 2006. [2] BARET F, GUYOT G, MAJOR D. TSAVI: a vegetation index which minimizes soil brightness effects on LAI and APAR estimation [C]// International Geoscience and Remote Sensing Symposium. 12th Canadian Symposium on Remote Sensing Geoscience and Remote Sensing Symposium. Vancouver: Institute of Electrical and Electronics Engineers, 1989: 1355 − 1358. [3] GEERKEN R, ZAITCHIK B, EVANS J P. Classifying rangeland vegetation type and coverage from NDVI time series using Fourier Filtered Cycle Similarity [J]. International Journal of Remote Sensing, 2005, 26(24): 5535 − 5554. [4] WARDLOW B D, EGBERT S L, KASTENS J H. Analysis of time-series MODIS 250 m vegetation index data for crop classification in the U. S. Central Great Plains [J]. Remote Sensing of Environment, 2007, 108(3): 290 − 310. [5] MASELLI F, CHIESI M. Integration of multisource NDVI data for the estimation of Mediterranean forest productivity [J]. International Journal of Remote Sensing, 2006, 27(1): 55 − 72. [6] PARUELO J M, EPSTEIN H E, LAUENROTH W K, et al. ANPP estimates from NDVI for the central grassland region of the United States [J]. Ecology, 1997, 78(3): 953 − 958. [7] ULLAH S, SI Yali, SCHLERF M, et al. Estimation of grassland biomass and nitrogen using MERIS data [J]. International Journal of Applied Earth Observation and Geoinformation, 2012, 19: 196 − 204. [8] 李晓松, 李增元, 高志海, 等. 基于NDVI与偏最小二乘回归的荒漠化地区植被覆盖度高光谱遥感估测[J]. 中国沙漠, 2011, 31(1): 162 − 167. LI Xiaosong, LI Zengyuan, GAO Zhihai, et al. Estimation of vegetation cover in desertified regions from Hyperion imageries using NDVI and partial least squares regression [J]. Journal of Desert Research, 2011, 31(1): 162 − 167. [9] 刘广峰, 吴波, 范文义, 等. 基于像元二分模型的沙漠化地区植被覆盖度提取——以毛乌素沙地为例[J]. 水土保持研究, 2007, 14(2): 268 − 271. LIU Guangfeng, WU Bo, FAN Wenyi, et al. Extraction of vegetation coverage in desertification regions based on the dimidiate pixel model: a case study in Maowusu Sandland [J]. Research of Soil and Water Conservation, 2007, 14(2): 268 − 271. [10] HILKER T, WULDER M A, COOPS N C, et al. Generation of dense time series synthetic Landsat data through data blending with MODIS using a spatial and temporal adaptive reflectance fusion model [J]. Remote Sensing of Environment, 2009, 113(9): 1988 − 1999. [11] CONGALTON R, ODERWALD R G, MEAD R. Assessing Landsat classification accuracy using discrete multivariate analysis statistical techniques [J]. Photogrammetric Engineering and Remote Sensing, 1983, 49(12): 1671 − 1678. [12] YUAN Fei, SAWAYA K E, LOEFFELHOLZ B C, et al. Land cover classification and change analysis of the Twin Cities (Minnesota) Metropolitan Area by multitemporal Landsat remote sensing [J]. Remote Sensing of Environment, 2005, 98(2): 317 − 328. [13] DUBE T, MUTANGA O. Evaluating the utility of the medium-spatial resolution Landsat 8 multispectral sensor in quantifying aboveground biomass in uMgeni catchment South Africa [J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2015, 101: 36 − 46. [14] SCHMIDT T, FORSTERET M, GÄRTNER P, et al. Prediction of NDVI for grassland habitats by fusing RapidEye and Landsat imagery [C]// Geoscience and Remote Sensing Society. 2015 8th International Workshop on the Analysis of Multitemporal Remote Sensing Images (Multi-Temp). Annecy: Institute of Electrical and Electronics Engineers, 2015: 1 − 4. [15] EMELYANOVA I V, MCVICAR T R, van NIEL T G, et al. Assessing the accuracy of blending Landsat-MODIS surface reflectances in two landscapes with contrasting spatial and temporal dynamics: a framework for algorithm selection [J]. Remote Sensing of Environment, 2013, 133(suppl C): 193 − 209. [16] REEVES M C, ZHAO Maosheng, RUNNING S W. Applying improved estimates of MODIS productivity to characterize grassland vegetation dynamics [J]. Rangeland Ecology &Management, 2006, 59(1): 1 − 10. [17] GAO Tian, XU Bin, YANG Xiuchun, et al. Using MODIS time series data to estimate aboveground biomass and its spatio-temporal variation in Inner Mongolia’s grassland between 2001 and 2011 [J]. International Journal of Remote Sensing, 2013, 34(21): 7796 − 7810. [18] GU Yingxin, WYLIE B K, BLISS N B. Mapping grassland productivity with 250 m eMODIS NDVI and SSURGO database over the Greater Platte River Basin, USA [J]. Ecological Indicators, 2013, 24(1): 31 − 36. [19] LI Fei, JIANG Lei, WANG Xufeng, et al. Estimating grassland aboveground biomass using multitemporal MODIS data in the West Songnen Plain, China [J/OL]. Journal of Applied Remote Sensing, 2013, 7(1): 073546[2022-06-01]. doi: 10.1117/1.JRS.7.073546. [20] JIN Yuanxiang, YANG Xiuchun, QIU Jianjun, et al. Remote sensing-based biomass estimation and its spatio-temporal variations in temperate grassland, northern China [J]. Remote Sensing, 2014, 6(2): 1496 − 1513. [21] BOYTE S P, WYLIE B K, RIGGE M B, et al. Fusing MODIS with Landsat 8 data to downscale weekly normalized difference vegetation index estimates for central Great Basin rangelands, USA [J]. GIScience &Remote Sensing, 2017, 55(3): 376 − 399. [22] GAO Feng, MASEK J, SCHWALLER M, et al. On the blending of the Landsat and MODIS surface reflectance: predicting daily Landsat surface reflectance [J]. IEEE Transactions on Geoscience and Remote Sensing, 2006, 44(8): 2207 − 2218. [23] 石月婵, 杨贵军, 李鑫川, 等. 融合多源遥感数据生成高时空分辨率数据的方法对比[J]. 红外与毫米波学报, 2015, 34(1): 92 − 99. SHI Yuechan, YANG Guijun, LI Xinchuan, et al. Intercomparison of the different fusion methods for generating high spatial-temporal resolution data [J]. Journal of Infrared and Millimeter Waves, 2015, 34(1): 92 − 99. [24] HOBYB A, RADGUI A, TAMTAOUI A, et al. Evaluation of spatiotemporal fusion methods for high resolution daily NDVI prediction [C]//GERAID S. 2016 5th International Conference on Multimedia Computing and Systems. Marrakech: Institute of Electrical and Electronics Engineers, 2016: 121 − 126. [25] ZHOU Junxiong, CHEN Jin, CHEN Xuehong, et al. Sensitivity of six typical spatiotemporal fusion methods to different influential factors: a comparative study for a normalized difference vegetation index time series reconstruction [J]. Remote Sensing of Environment, 2021, 252: 112 − 130. [26] WANG Qunming, ATKINSON P M. Spatio-temporal fusion for daily Sentinel-2 images [J]. Remote Sensing of Environment, 2018, 204: 31 − 42. [27] JIANG Chong, ZHANG Linbo. Climate change and its impact on the eco-environment of the Three-Rivers Headwater Region on the Tibetan Plateau, China [J]. International Journal of Environmental Research and Public Health, 2015, 12(10): 12057 − 12081. [28] ZENG Na, REN Xiaoli, HE Hongli, et al. Estimating the grassland aboveground biomass in the Three-River Headwater Region of China using machine learning and Bayesian model averaging [J/OL]. Environmental Research Letters, 2021, 16: 114020[2022-06-01]. doi: 10.1088/1748-9326/ac2e85. [29] 曾纳, 任小丽, 何洪林, 等. 基于神经网络的三江源区草地地上生物量估算[J]. 环境科学研究, 2017, 30(1): 59 − 66. ZENG Na, REN Xiaoli, HE Honglin, et al. Aboveground biomass of grasslands in the Three-River Headwaters Region based on neural network [J]. Environmental Research Letters, 2017, 30(1): 59 − 66. [30] 易湘生, 尹衍雨, 李国盛, 等. 青海三江源地区近50年来的气温变化[J]. 地理学报, 2011, 66(11): 1451 − 1465. YI Xiangsheng, YIN Yanyu, LI Guosheng, et al. Temperature variation in recent 50 years in the Three-River Headwaters Region of Qinghai Province [J]. Acta Geographica Sinica, 2011, 66(11): 1451 − 1465. [31] 刘纪远, 徐新良, 邵全琴. 近30年来青海三江源地区草地退化的时空特征[J]. 地理学报, 2008, 18(4): 364 − 376. LIU Jiyuan, XU Xinliang, SHAO Quanqin. Grassland degradation in the “Three-River Headwaters” region, Qinghai Province [J]. Acta Geographica Sinica, 2008, 18(4): 364 − 376. [32] 吴炳方, 苑全治, 颜长珍, 等. 21世纪前十年的中国土地覆盖变化[J]. 第四纪研究, 2014, 34(4): 723 − 731. WU Bingfang, YUAN Quanzhi, YAN Changzhen, et al. Land cover changes of China from 2000 to 2010 [J]. Quaternary Sciences, 2014, 34(4): 723 − 731. [33] ZHU Xiaolin, CHEN Jin, GAO Feng, et al. An enhanced spatial and temporal adaptive reflectance fusion model for complex heterogeneous regions [J]. Remote Sensing of Environment, 2010, 114(11): 2610 − 2623. [34] GU Yingxin, WYLIE B K. Downscaling 250 m MODIS growing season NDVI based on multiple-date Landsat images and data mining approaches [J]. Remote Sensing, 2015, 7(4): 3489 − 3506. [35] LIU Maolin, KE Yinghai, YIN Qi, et al. Comparison of five spatio-temporal satellite image fusion models over landscapes with various spatial heterogeneity and temporal variation [J/OL]. Remote Sensing, 2019, 11(22): 2612[2022-06-01]. doi: 10.3390/rs11222612. [36] 张小利, 李雄飞, 李军. 融合图像质量评价指标的相关性分析及性能评估[J]. 自动化学报, 2014, 40(2): 306 − 315. ZHANG Xiaoli, LI Xiongfei, LI Jun. Validation and correlation analysis of metrics for evaluating performance of image fusion [J]. Acta Automatica Sinica, 2014, 40(2): 306 − 315. [37] ZHANG Binghua, ZHANG Li, XIE Dong, et al. Application of synthetic NDVI time series blended from Landsat and MODIS data for grassland biomass estimation [J/OL]. Remote Sensing, 2016, 8(1): 10[2022-06-01]. doi: 10.3390/rs8010010. [38] LI Aihua, BO Yanchen, ZHU Yuxin, et al. Blending multi-resolution satellite sea surface temperature (SST) products using Bayesian maximum entropy method [J]. Remote Sensing of Environment, 2013, 135: 52 − 63. -
链接本文:
https://zlxb.zafu.edu.cn/article/doi/10.11833/j.issn.2095-0756.20220381