Using land surface phenology for spatio-temporal. temporal mining of image time series: a manifesto

Transcription

1 Using land surface phenology for spatio-temporal temporal mining of image time series: a manifesto Geoffrey M. Henebry Kirsten M. de Beurs Geographic Information Science Center of Excellence (GIScCE( GIScCE) South Dakota State University {Geoffrey.Henebry sdstate.edu GIScCE

2 Outline of Talk 0. Synoptic ecology: remote sensing of ecological change 1. What are the appropriate units of analysis? 2. What constitutes an appropriate baseline? 3. How do we conduct change analysis? 4. Concluding thoughts

3 Biogeography (decades to millenia to eons) Synoptic Ecology (weeks to years to decades) Human Land Use Meteorology (seconds to hours to weeks) Climate Climate Landscape Vegetation Vegetation 0. Why Synoptic Ecology? Weather Synoptic ecology focuses on the interface between land surface & atmosphere, where human activity greatly affects the biogeophysical & socioeconomic patterns and processes. Science requires the ability to distinguish unusual change from expected variation. This assumes a good notion of what is expected and thereby provides a foundation for modeling. Vegetation Long range objective: Development of an operational environmental monitoring & forecasting system for the land surface dynamics that interacts with other environmental monitoring & forecasting systems, e.g., the global weather network. Weather

4 1. What are the Units of Analysis? A DEEP PROBLEM What The tools are and the techniques appropriate commonly units used of to analysis represent and (UoA( manipulate geospatial data carry strong assumptions as to what constitutes the units of analysis. UoA) ) for image time series of biogeophysical fields localized with spatial and temporal coordinates? The phrase units of analysis refers not to units of specific measurement systems but rather to the conceptual entities that are subject to measurement and analysis. Individual pixels are neither Geospatial entities include axiomatic geometric objects (e.g., points, lines, polygons, polyhedra) appropriate that are located within nor a sufficient! spatial reference system. But geospatial entities can also be synthetic geometric objects derived from sensor systems, such as a spatio-temporal trajectory representing telemetered animal movements or an array of pixels portraying hyperspectral radiance upwelling from a landscape. Fisher, P The pixel: a snare and a delusion. International Journal of Remote Sensing 18: Cracknell, Henebry, G.M., A.P. and J. W. Review Merchant. article Synergy Geospatial in remote data sensing time: limits what s and in a prospects pixel? International for predicting Journal species of Remote occurrences. Sensing 19: In Predicting Species Occurrences: Issues of Scale and Accuracy, (J.M. Scott, P. J. Heglund, M.L. Morrison, J.B. Haufler, M.G. Raphael, W.A. Wall, and F.B. Samson, editors). Island Press, Covello, CA. Chapter 23. pp

5 The UoA question involves resolution & extent of each dimension of measurement AND the characteristic extents of the thing(s) ) observed. Let R = scene object extent / sensor resolution. If R << 1, then multiple scene objects can fall within one sensor grain. Gradients in space. Fluctuations in time. Summarized by measures of dependence, e.g., correlation length, range. If R >> 1, then multiple sensor grains can portray a scene object. Contrasts in space. Seasonality in time. Summarized by measures of heterogeneity, e.g., diversity/entropy. Inspired by: Strahler, A.H., C.E. Woodcock, & J.A. Smith On the nature of models in remote sensing. Remote Sensing of Environment 20:

6 There are no natural a priori spatial units! We impose units by our observational processes. Thus, delineations between patches are arbitrary and may be imprecise in location, transitory in duration, and irrelevant to underlying processes of interest. Further, there is no a priori ordering of the directionality of causation in space comparable to the arrow of time. While topological relationships indicate who is the neighbor of whom, additional information is required to know who are the effective neighbors. This requires the user to inform the geospatial database about the flows of influence among spatially ordered data. Different processes can have different effective neighborhoods at different scales.

7 Pattern searching is not the same as hypothesis testing because there is no relevant null hypothesis. This point was lost on the original quantitative geographers [during the 1970 s]. [They] failed to develop image a statistical time theory of series spatial analysis are as distinct not from the providing examples of statistical methods being applied to spatial data in pictures search for largely themselves, aspatial patterns. The danger, but now is the that the same mistake will be repeated 20 years later in the GIS era by a failure to appreciate that spatial patterns are themselves geographic objects that can be recognized and extracted from spatial databases. What is of scientific interest in dynamic of pattern and process that sequences of pictures portray. From: Openshaw, S A concepts-rich approach to spatial analysis, theory generation, and scientific discovery in GIS using massively parallel computing. In: (M. Worboys, ed.) Innovations in GIS. London: Taylor and Francis. pp

8 2. What constitutes an appropriate baseline? Statistical modeling of complex spatio temporal temporal data through local filters on neighbors (AR) and noise (MA), (MA), e.g., AR(I)MA, CAR, SAR, wavelets, harmonic/fourier analysis, Kalman filters, etc. Climatological approaches via the moments of the distribution of accumulated observations. Use the power of recurrent observation for identifying anomalies, the unusual, and the unique. Building empirical expectations sets the stage for change analysis & forecasting.

9 SEEKING BASELINES AN AN ANALOGY: Consider sparsely sampling a movie by individual frames or even frame sequences. One level of analysis might aim at reconstructing motion, but a more sophisticated analysis would aim at reconstructing the plot. Intelligent (and informed) knowledge discovery in scientific databases must aim at * reconstructing plots, * comparing plots, * identifying unusual plots as well as interesting deviations from typical plots.

10 Some relevant ecological plots include: Succession in ecological communities/ecosystem structure Growth and development of urban areas Disaster recovery Invasive species/disease outbreak & spread Land surface phenology

11 Land Surface Phenology: the what and the why Land Surface Phenology is defined as the spatio-temporal patterns of the vegetated land surface as observed by synoptic sensors at spatial resolutions and extents relevant to meteorological processes in the atmospheric boundary layer. Land Surface Phenology deals with mixtures of land covers; it is distinct from the phenology of particular species. Linked to seasonality of aboveground net primary production (ANPP). Need to understand the dynamics of Land Surface Phenology to monitor & model carbon, water, energy exchanges in the biosphere.

12 We can observe Land Surface Phenology from orbital platforms by sensing reflected solar radiation MODIS 1km R=08MAY2004 G=27JUL2004 B=08MAR2004

13 We can observe Land Surface Phenology from orbital platforms by sensing emitted terrestrial radiation 25 km Vegetation Water Content 2005 descending mode R=CV G=mean B=skew

14 We can observe Land Surface Phenology from orbital platforms by sensing backscattered anthropogenic radiation 25 m ST2 composite R=05MAY1999 G=PC1 of 6 scenes B=26AUG1999

15 3. How do we conduct change analysis? 1. Change Detection perceiving the differences 2. Change Quantification measuring the magnitudes of differences 3. Change Assessment determining the significances of differences 4. Change Attribution identifying/inferring the proximate causes 5. Change Consequences ancillary data, modeling, domain expertise

16 In a recently concluded NASA LCLUC project, we investigated whether the changes in the agricultural sector consequent to the collapse of the Soviet Union had led to changes in land cover and/or land use that would be sufficiently widespread to be observable at spatio-temporal scales that could affect exchanges of water and energy between the land surface and the atmospheric boundary layer. We focused our analysis on the onset of spring because the widespread commencement of vegetation growth causes substantial shifts in the surface energy balance.

17 To model the spring green-up, we used two freely available and widely used time series: (1) Pathfinder AVHRR Land (PAL) NDVI dataset and (2) near-surface air temperature from the NCEP/NCAR Reanalysis. Big spatio-temporal scale differences: 10-day maximum value composite at 8km resolution (PAL NDVI) vs. daily temperature extrema on a 2 o global grid (NCEP/NCAR). We calculated accumulated growing degree-days (AGDD), which is a kind of thermal time, and we modeled LSP by linking NDVI to AGDD using two different forms of the relationship: (1) a linear quadratic model for herbaceous vegetation in croplands & grasslands (2) a nonlinear spherical model for the initial green-up and plateau exhibited by ecoregions dominated by woody vegetation.

18 WWF ecoregion: : Kazakh Steppe

19 WWF ecoregion: : Kazakh Steppe &

20 WWF ecoregion: : Kazakh Steppe de Beurs, K.M., and G.M. Henebry. 2005a. A statistical framework for the analysis of long image time series. International Journal of Remote Sensing 26(8):

21 Change in LSP is not uniform across KZ (de Beurs and Henebry 2005a) NDVI NDVI NDVI AGDD AGDD AGDD

22 WWF ecoregion: : Northeast Siberian Taiga

23 WWF ecoregion: : Northeast Siberian Taiga &

24 WWF ecoregion: : Northeast Siberian Taiga de Beurs, K.M., and G.M. Henebry. 2005b. Land surface phenology and temperature variation in the IGBP highlatitude transects. Global Change Biology 11(5):

25 Both LSP models fit well in some regions and neither does in others. Applying the better fit to NCEP grid cells across the northern hemisphere, we can map out four ecologically interpretable metrics that can be derived from the LSP model parameters: (1) the NDVI at the onset of the observing season; (2) the seasonal peak NDVI; Density sliced images of PC1s from 9 y ( & ) (3) the quantity of AGDD needed to reach the peak; and (4) the seasonal dynamic range of NDVI.

26 4. Concluding thoughts We used these LSP models as biometeorological filters on the two image time series to reveal significant changes in spring s greening in the wake of the Soviet Union s collapse. In challenging the image time series with specific functional models informed by our ecological understanding, we enhance our ability to detect where we understand the data and where the models break down.

27 Take Home Points 1. What are the appropriate units of analysis? It is very much depends on the question at hand and the available measurements, but they re unlikely to be pixels. 2. What constitutes appropriate baselines? Climate analogy: expectations based on prior observations. 3. How do we conduct change analysis? i. Detection ii. Quantification iii. Assessment iv. Attribution v. Consequences

28 Acknowledgments Recent Collaborators Ms. Marcela Doubková,, M.A. student in Geography, University of Nebraska-Lincoln Prof. William Capehart,, South Dakota School of Mines & Technology Prof. Anatoly Gitelson, University of Nebraska-Lincoln Dr. William Hargrove, Oak Ridge National Laboratory Dr. Forrest Hoffman, Oak Ridge National Laboratory Prof. Elena Lioubimtseva,, Grand Valley State University Prof. Walter Schacht, University of Nebraska-Lincoln Geographic Information Science Center of Excellence South Dakota State University GIScCE

29 Elaine s discussion questions: Q1. What data mining or statistical methods were used? A1. Parametric modeling, significance testing, PCA, clustering Q2. How were the techniques developed? a) from scratch in response to a specific Earth science question, or b) modified from another application or research project? A2. Standard methods structured according to a specific question. Q3. What is the importance of the science question? A3. LSP is a key linkage between terrestrial vegetation and boundary layer processes. Q4. What scientific results were obtained that would have been difficult or impossible without data mining and/or statistics? A4. Assessment of the significance of detected changes in LSPs. Q5. What were the obstacles? A5. Little relevant prior literature; differences among sensors; paucity of met data at right scales. Q6. Did this work result in a geoscience publication? If not, why not? A.6 Yes: 1. de Beurs, K.M., and G.M. Henebry. 2004a. Land surface phenology, climatic variation, and institutional change: Analyzing agricultural land cover change in Kazakhstan. Remote Sensing of Environment 89: de Beurs, K.M., and G.M. Henebry. 2004b. Trend analysis of the Pathfinder AVHRR Land (PAL) NDVI data for the deserts of Central Asia. IEEE Geoscience and Remote Sensing Letters 1(4): de Beurs, K.M., and G.M. Henebry. 2005a. A statistical framework for the analysis of long image time series. International Journal of Remote Sensing 26(8): de Beurs, K.M., and G.M. Henebry. 2005b. Land surface phenology and temperature variation in the IGBP high-latitude transects. Global Change Biology 11(5): de Beurs, K.M., and G.M. Henebry. Human impact on land cover dynamics in the extensive agricultural regions of the Northern Hemisphere. Journal of Geophysical Research-Biogeosciences, in revision following review.