GOPHER: GLOBAL OBSERVATION OF PLANETARY HEALTH AND ECOSYSTEMRESOURCES

Ashish Garg, Varun Mithal, Yashu Chamber,Ivan Brugere, Vijay Chaudhari,

Marc Dunham, Vikrant Krishna,Sairam Krishnamurthy, Sruthi Vangala,

Shyam Boriah, Michael Steinbach, Vipin Kumar

University of Minnesota

Albert Cho, JD Stanley, Teji Abraham,Juan Carlos Castilla-Rubio

Planetary Skin Institute

Christopher Potter, Steven Klooster

NASA Ames Research Center

Index Terms— Forestry, Remote monitoring, Data min-ing, Time series analysis.

1. INTRODUCTION

Land cover change, especially deforestation, is a priority issuefor policymakers at the local, national and international scale.Deforestation’s contribution of up to 20% of global green-house gas emissions is already well known; the loss of biodi-versity from land conversion is also well established [7, 13].Policymakers at the UN Framework Convention on ClimateChange negotiations are addressing land use change by de-veloping a framework for Reducing Emissions from Defor-estation and Degradation (REDD). In parallel, a number oftechnical groups have been working to identify strategies forreliably monitoring, reporting and verifying land use changeand emissions [3, 2].

However, while a number of detailed and comprehensiveoptions have been examined to implement a long-term globalmonitoring capability, in the medium term most tropical de-veloping countries do not have access to reliable methods fortracking land cover change in their forests. Many promis-ing initiatives have demonstrated the ability to use satelliteimagery to track deforestation through image comparisons,but these approaches are difficult to scale globally as they aretime-intensive and require skilled operators.

Given the urgency and importance of the situation, in-expensive and rapidly scalable solutions are needed, even ifthey are imperfect or operate at coarser resolutions than anideal long-term solution. GOPHER (Global Observation ofPlanetary Health and Ecosystem Resources) is a collection ofdata mining algorithms for detecting global land use and landcover change that builds on a decade of research on spatio-temporal data mining at the University of Minnesota; the GO-PHER approach to analyzing remote sensing imagery pro-vides such a solution by providing rapid, inexpensive, robust,scalable, and precise detection of land use change. GOPHERalgorithms are able to detect historical changes with high pre-

cision, and more recent changes with reasonable precisionin as little as 8 weeks after changes occur (this is due to acombination of remote sensing data availability and model-ing needs).

This paper outlines the GOPHER approach and providessome illustrative results to demonstrate its utility.

2. GOPHER

Most previous change detection studies have relied on exam-ining differences between two or more satellite images ac-quired on different dates [6]. The key innovation of our ap-proach is to use repeated satellite observations to view everypixel on the planet as a time series rather than a static image.By tracking the value of a specific indicator or set of indi-cators over time, the problem of land use change detectionceases to be one of point-to-point comparison, and becomesa time series analysis problem amenable to advanced compu-tational analysis. In the case of land use change detection,data mining algorithms are applied to the evolution of vege-tation reflectance indicators for each pixel to identify signifi-cant changes in the vegetation signal over time. This sectionbriefly describes the overall approach, including data sources,algorithms, and the methodology used to detect disturbanceson a global scale.

2.1. Data

In principle, the suite of GOPHER algorithms can be ap-plied to any geospatial dataset that features regular, repeatedobservations, consistent image registration and well-definedcomposite indicators of vegetation. Currently, we use theEnhanced Vegetation Index (EVI), a data product based onmeasurements taken by the Moderate Resolution ImagingSpectroradiometer (MODIS) sensor on NASA’s Terra satel-lite and distributed through the Land Processes DistributedActive Archive Center (LP DAAC). EVI essentially measuresthe “greenness” signal (area-averaged canopy photosynthetic

capacity) as a proxy for the amount of vegetation at a par-ticular location. MODIS data have been used to generatea continuous record of the EVI index at spatial resolutionsof 250 meters and 1 km from February 2000 to the present.This index is generated at a temporal frequency of 16 days:each instance in the product is composited using the highestquality data from 16 daily raw observations.

GOPHER algorithms exploit several key advantages ofthe MODIS EVI product. First, while MODIS-based productshave a maximum resolution of 250 m, they have spatial andtemporal advantages over other data sources. For example,MODIS provides wall-to-wall global coverage free of costwith a short latency of 2–4 weeks. The daily revisit coverageprovides for the highest likelihood of data observations freeof atmospheric interference such as clouds and aerosols. Onthe other hand, while Landsat and radar-based platforms (e.g.ALOS PALSAR) have a higher spatial resolution, they havecrucial limitations preventing their use at global scale. Inparticular, the revisit time of Landsat is more than two weeks,which means the likelihood of obtaining cloud-free images issignificantly lower than for MODIS. This issue is especiallyacute in the tropics where disturbed vegetation may haveregrown by the time the next cloud-free image has becomeavailable. Finally, while radar-based products are cloud pene-trating, data sets with global wall-to-wall coverage are not yetwidely available. Furthermore, relative to NDVI (NormalizedDifference Vegetation Index) and FPAR (Fraction of Pho-tosynthetically Active Radiation), two other MODIS-basedproducts, EVI displays better performance in the presenceof atmospheric aerosols, and improved sensitivity in highbiomass cover areas.

Many other datasets could have been used in this analy-sis, including Landsat (at 30 m resolution). Landsat data wasnot used for this analysis, as there are image registration chal-lenges that make it difficult to track values for a specific pixelconsistently over time. In the future we hope to apply ourapproach to other sources of time series data, including 10 mPALSAR data from JAXA, the Japanese space agency. Thereis potential for combining all of these data sets and this is adirection we intend to explore in the future.

2.2. Algorithms

Change detection is the process of identifying changes in thetype and/or human use of the Earth’s land cover. A largebody of research has developed approaches to change detec-tion using remotely sensed image data [6, 10, 11]. As men-tioned before, most of these approaches perform change de-tection by comparing two or more images of different dates,or snapshots, to determine the land cover differences betweenthem. Because these approaches are difficult to scale glob-ally, most studies have focused on relatively small areas orhave described changes in very specific categories of interest.Recent work has shown the value of high temporal resolution

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

Time(18!Feb!2000 to 19!Dec!2010)

01!

Jan!

2001

01!

Jan!

2002

01!

Jan!

2003

01!

Jan!

2004

01!

Jan!

2005

01!

Jan!

2006

01!

Jan!

2007

01!

Jan!

2008

01!

Jan!

2009

01!

Jan!

2010

EV

I (!

10

4)

Fig. 1. Sudden drop.

(bi-weekly) time series data for land cover change detectionin applications where high spatial resolution images are im-practical, e.g., large study areas [11, 4, 12, 8].

Time series based change detection has significant ad-vantages over the comparison of snapshot images of selecteddates because it takes into account information about the tem-poral dynamics of land cover changes. Detection of changesis based on the pattern of spectral response of the landscapeover time rather than the differences between two or moreimages collected on different dates. Therefore additional pa-rameters such as the rate of the change (e.g. a sudden forestfire vs. gradual logging), the extent, and pattern of regrowthcan be derived.

The objective of the GOPHER algorithms is to detectwhen the characteristics of the EVI time series has changed.Characteristics can include a number of properties: mean,variance, trend, shape, seasonal characteristics, etc. Thus, fora given pixel (or location on the earth) we have as input anEVI time series; change detection algorithms then processthe time series and assign a score based on the characteristicsthat have changed. This change score denotes the extent towhich the EVI time series for the pixel has changed. In orderto do this effectively, the algorithms must overcome manychallenges: seasonality (which generates cyclical fluctuationsin the value of indicators), inherent variability, missing data(e.g. from cloud cover), or poor quality data. Moreover, theymust be capable of detecting many different types of change,including deforestation, gradual degradation, reforestation, orshifts from one type of production cycle to another. The GO-PHER suite of algorithms currently includes three categoriesof algorithms, comprising: sudden drops, gradual changes,and segmentation. Each is described below.

Sudden drop. The sudden drop algorithm identifies large andunexpected declines in greenness, such as would be exhibitedby fire or mechanized land clearance. To detect these changes,this algorithm constructs a model of EVI behavior from olderEVI observations and then uses this model to predict EVIfor more recent EVI time steps. If the predictions are suffi-

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

Time(18!Feb!2000 to 01!Jan!2011)

01!

Jan!

2001

01!

Jan!

2002

01!

Jan!

2003

01!

Jan!

2004

01!

Jan!

2005

01!

Jan!

2006

01!

Jan!

2007

01!

Jan!

2008

01!

Jan!

2009

01!

Jan!

2010

01!

Jan!

2011

EV

I (!

10

4)

Fig. 2. Gradual increase.

ciently different from what was observed at these time points,(i.e. it is not plausible the observed values could be gener-ated by the model), these time points are classified as changepoints. More specifically, each time point is assigned a scorethat measures the magnitude of the discrepancy between whatthe model predicts for EVI and what was observed. If therewas a sudden drop in the vegetation (as in the case of a fireor deforestation) the change score is high for the time stampin which the sudden drop occurs. For some pixels, due to thetype of vegetation, noise, or atmospheric conditions, vegeta-tion response drops abruptly without an actual change occur-ring on the ground. To avoid detecting false alarms in suchtime series, the algorithm normalizes the change score withthe expected natural variability of the vegetation response inprevious years. The natural variability of a EVI time seriesis computed by looking at the historical differences of EVIvalues between pairwise yearly segments in that time series.Figure 1 shows an example of a sudden drop.

Gradual increase or decrease. Some land cover changes oc-cur gradually, spanning several years. The gradual decreasealgorithm is designed to detect slow degradation of vegetationcover over multiple years. The main intuition behind the algo-rithm is to find time series for which the later years have lowerEVI values than previous years. Time series that exhibit thisbehavior for a larger number of consecutive years are given ahigher score by the algorithm (e.g. see Figure 3). A similarapproach is used to detect time series that indicate vegetationgrowth. In this case we look for time series for which the fol-lowing years have higher EVI values than previous years (e.g.see Figure 2).

Segmentation. The goal of segmentation is to divide a timeseries into homogeneous parts that are different from eachother. Segmentation can be used to detect land cover or landuse type conversions where both the pre-change and post-change vegetation phenology are stable but different fromeach other. The segmentation-based algorithm computes thecohesion within each segment and the separation (distance)between segments. The measures for cohesion of segments

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

Time(18!Feb!2000 to 19!Dec!2010)

01

!Ja

n!

20

01

01

!Ja

n!

20

02

01

!Ja

n!

20

03

01

!Ja

n!

20

04

01

!Ja

n!

20

05

01

!Ja

n!

20

06

01

!Ja

n!

20

07

01

!Ja

n!

20

08

01

!Ja

n!

20

09

01

!Ja

n!

20

10

EV

I (!

10

4)

Fig. 3. Gradual decrease.

and separation are then combined to determine the appropri-ate change score and the time of change for the time series ofeach pixel. Figure 4 shows an example of a time series foundusing the segmentation algorithm.

2.3. Results

The GOPHER algorithms detected over a million 1 kmchanged pixels globally in the last decade with high confi-dence using the EVI data set. These changes are available forinteractive exploration via the ALERTS (Automated Land-change Evaluation, Reporting and Tracking System) platform(www.ourplanetaryskin.org). Most of the changesfound correspond to fires, large scale deforestation, floodsand other variety of forest changes all across the globe. Theresults were evaluated using validation data sets obtainedfrom various government agencies and NGOs that maintainrecords of forest disturbances in different parts of the world.Validation was performed against PRODES data maintainedby INPE in Brazil [5], SarVision in South East Asia [9], andforest fire data from governments in Canada [14] and Cali-

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

Time(18!Feb!2000 to 19!Dec!2010)

01

!Ja

n!

20

01

01

!Ja

n!

20

02

01

!Ja

n!

20

03

01

!Ja

n!

20

04

01

!Ja

n!

20

05

01

!Ja

n!

20

06

01

!Ja

n!

20

07

01

!Ja

n!

20

08

01

!Ja

n!

20

09

01

!Ja

n!

20

10

EV

I (!

10

4)

Fig. 4. This figure shows the EVI time series that correspondsto a location in Zimbabwe where a change in crop type haslikely occurred in 2002.

Fig. 5. Overlap of changed pixels identified by GOPHER al-gorithms with SarVision validation data (red polygons) in In-donesia.

Fig. 6. Overlap of changed pixels identified by GOPHERalgorithms with PRODES validation data (blue polygons) inBrazil.

fornia [1]. The analysis indicated that there was considerableoverlap between the changes identified by our algorithmsand validation data, and that GOPHER algorithms identifiedchanges that were either missed by other sources or werenot in the scope of their analysis. Figures 5 and 6 show theoverlap of GOPHER events shown as red dots with the defor-estation polygons from SarVision in Indonesia and PRODESin Brazil respectively.

3. FUTURE WORK

GOPHER algorithms face significant challenges due to thepresence of noise in the EVI data set. We are currently in-vestigating a number of time series reconstruction techniquesto reduce noise in the data sets. A key aspect of this work ison smoothing time series without affecting large shifts due toactual land cover change. We are also extending the GO-PHER change detection algorithms to work with a varietyof land cover types like savannas and shrubs. These landcover types undergo changes of interest, but the high vari-ability in their greenness due to inter-annual changes in pre-cipitation and temperature creates additional challenges. Fi-nally, it is of considerable advantage for the end-user of a landcover change visualization system to see a summarized view

of land cover changes. To this end, we are developing spatial-temporal clustering algorithms that group events and presentthem in a more compact and informative view.

References[1] California Department of Forestry and Fire Protection, Fire and

Resource Assessment Program.http://frap.fire.ca.gov.

[2] Terrestrial Carbon Group Policy Briefs.http://www.terrestrialcarbon.org/Publications.aspx.

[3] Monitoring forests: Seeing the world for the trees. TheEconomist, December 16 2010.

[4] S. Boriah, V. Kumar, M. Steinbach, C. Potter, and S. Klooster.Land cover change detection: A case study. In KDD ’08: Pro-ceedings of the 14th ACM SIGKDD International Conferenceon Knowledge Discovery and Data Mining, pages 857–865,2008.

[5] G. Camara, D. d. M. Valeriano, and J. V. Soares. Metodologiapara o calculo da taxa anual de desmatamento na Amazonialegal. Sao Jose dos Campos, INPE, 2006.

[6] P. Coppin, I. Jonckheere, K. Nackaerts, B. Muys, and E. Lam-bin. Digital change detection methods in ecosystem moni-toring: a review. International Journal of Remote Sensing,25(9):1565–1596, 2004.

[7] R. E. Gullison, P. C. Frumhoff, J. G. Canadell, C. B.Field, D. C. Nepstad, K. Hayhoe, R. Avissar, L. M. Curran,P. Friedlingstein, C. D. Jones, and C. Nobre. Tropical forestsand climate policy. Science, 316(5827):985–986, 2007.

[8] D. Hammer, R. Kraft, and D. Wheeler. FORMA: Forestmonitoring for action—rapid identification of pan-tropical de-forestation using moderate-resolution remotely sensed data.Working Paper 192, Center for Global Development, 2009.

[9] D. Hoekman, V. Schut, M. Vissers, and M. Nugroho.Fast medium resolution Indonesian forest monitoring system(World Bank contract: Towards Outcome-based Regulation ofDecentralised Forest Management through Tracking AnnualChanges in Actual Forest Cover). SarVision Report, Wagenin-gen, 2004.

[10] D. Lu, P. Mausel, E. Brondızio, and E. Moran. Change de-tection techniques. International Journal of Remote Sensing,25(12):2365–2401, 2003.

[11] R. S. Lunetta, J. F. Knight, J. Ediriwickrema, J. G. Lyon,and L. D. Worthy. Land-cover change detection using multi-temporal MODIS NDVI data. Remote Sensing of Environment,105(2):142–154, 2006.

[12] V. Mithal, A. Garg, S. Boriah, M. Steinbach, V. Kumar, C. Pot-ter, S. Klooster, and J. C. Castilla-Rubio. Monitoring globalforest cover using data mining. ACM Transactions on Intelli-gent Systems and Technology, to appear, 2011.

[13] N. Ramankutty, H. K. Gibbs, F. Achard, R. Defries, J. A. Fo-ley, and R. A. Houghton. Challenges to estimating carbonemissions from tropical deforestation. Global Change Biology,13:51–66, January 2007.

[14] B. J. Stocks, J. A. Mason, J. B. Todd, E. M. Bosch, B. M.Wotton, B. D. Amiro, M. D. Flannigan, K. G. Hirsch, K. A.Logan, D. L. Martell, and W. R. Skinner. Large forest fires inCanada, 1959–1997. J. Geophys. Res., 107, 2002.