Evaluation of Landsat and MODIS data fusion products for analysis of dryland forest phenology
Highlights
► Dryland forest areas present spatial and temporal challenges for phenology studies. ► Fusion of high-spatial/high-temporal resolution data may yield appropriate data. ► Different MODIS datasets offer various benefits when fused with Landsat data. ► BRDF-adjusted datasets yield superior fused imagery for phenological research.
Introduction
Arid and semi-arid ecosystems (i.e., “drylands”) are characterized by their scarcity of water; they are typically defined as areas in which annual water loss through evapotranspiration exceeds annual moisture inputs from precipitation (Bailey, 1998, Neary et al., 2002, Safriel et al., 2005). These ecosystems are not exclusively composed of isolated and barren desert landscapes. Of the 41% of the Earth's terrestrial surface that comprises drylands, 18% is covered by forest and woodland systems (Safriel et al., 2005). Dryland forests occupy large swaths of the western conterminous United States; more than 24 million hectares of dryland forest are found in the interior western states (Arizona, Colorado, Idaho, Montana, Nevada, New Mexico, Utah, and Wyoming) (Smith et al., 2009). Extensive tracts of dryland forest are also found in western Texas (Coulston et al., 2010), the southern Black Hills of South Dakota (Shepperd and Battaglia, 2002) and along the central and eastern portions of California, Oregon, and Washington (Bolsinger and van Hees, 1989).
The behavior of dryland forests under altered climatic conditions is a matter of environmental and societal importance given the critical ecosystem services these forests supply to human populations (Safriel et al., 2005). Diverse forest ecosystems may have already begun to respond to the elevated temperatures and/or water stress associated with climate change through an increase in mortality events across the globe (Allen et al., 2010). The sensitivity of drylands to climate variability (Hufkens et al., 2008) may exacerbate the reactions of dryland forests to future climate alterations. Dryland vegetation is acutely vulnerable to changes in rainfall timing or amount (Brown et al., 1997), while drought-induced mortality in dryland tree species has been observed to increase in the presence of higher temperatures (Adams et al., 2009, Breshears et al., 2005). Given that most climate models predict a high likelihood of reduced water resource availability across the western U.S. (IPCC, 2007) and an excess of 2 °C warming over the next century (Christensen et al., 2007), regional die-offs of dryland vegetation may increase in both extent and frequency in the future (Adams et al., 2009).
These anticipated but indefinite changes to a valuable ecological resource underscore the importance of characterizing current dryland forest behavior in order to estimate future response under climate change (Hicke et al., 2007). The use of remote sensing data to investigate spatial and temporal changes in vegetated surfaces, or “land surface phenology” (de Beurs and Henebry, 2004, Friedl et al., 2006, Morisette et al., 2009), is ideal for this purpose. Alterations in phenological metrics, such as the dates of spring green-up, can be used to provide an independent measure of the response of ecosystems to climate variability (White et al., 2009). To be most effective, the spatial and temporal resolution of the remote sensing data must be attuned to the characteristic phenological signature of the landscape under consideration. Drylands are subject to complex precipitation regimes in terms of timing, magnitude, and variation (Loik et al., 2004), all of which can contribute to a mosaic of plant phenology states at patch, landscape, and regional scales (Asner et al., 2003, Reid et al., 1999, Zhang et al., 2003). These patches can be spatially scattered across the landscape; the phenological changes that result from the characteristically patchy mortality events found in drylands (Reid et al., 1999) can occur throughout the full range of a species' distribution, including in areas of optimal habitat where higher tree density results in increased competition for available resources (Allen et al., 2010, Breshears et al., 2005).
By virtue of its integration of spectral responses over broad areas, a coarse-resolution satellite sensor collects data that subsumes fine-scale changes in vegetation phenology patterns. A remote sensing analysis using 1 km2 resolution data, for instance, determined that in dryland forests with moderate tree mortality, herbaceous response could partially compensate for smaller-scale tree mortality events (Breshears et al., 2005). Phenological studies of dryland areas may benefit from a simultaneous analysis of vegetation dynamics at scales ranging from stand to regional. Restricting an analysis to regional views may miss important dynamics on smaller scales, while analyses that focus exclusively on smaller-scale responses may lead to mistaken inferences about system-wide trends. Baseline information about the reaction of dryland forests to climate conditions across multiple scales may allow researchers and forest managers to distinguish between isolated events of natural variability and precursors of broader ecosystem distress.
One of the most commonly used datasets in land surface phenology studies is collected by the MODerate Resolution Imaging Spectroradiometer (MODIS) sensors on the Terra and Aqua satellites. The MODIS sensors collect data over a 110° field of view at moderate spatial resolutions (250–500 m for land cover bands) on a daily basis (Wolfe et al., 1998). The frequent availability of 500 m surface reflectance data is well-suited for tracking the timing of vegetation dynamics events over landscapes, but does not easily allow an examination of phenological dynamics on a sub-pixel level. This spatial constraint is problematic for fine-scale phenological analyses of the western U.S. drylands. The data source that is most suited for the examination of smaller-scale changes within landscape extents is delivered by the Landsat satellites. Landsat sensors collect data over a 15° field of view at a 30 m spatial resolution, which is suitable for individual forest stand management and analysis (Goward et al., 2008). However, the imagery is limited by the satellite revisit time of 16 days, which is frequently lengthened due to cloud cover (Ju and Roy, 2008). Even projects such as the Web-enabled Landsat Data (WELD), which compiles Landsat Enhanced Thematic Mapper (ETM +) data on a monthly, seasonal, and yearly basis over the conterminous U.S. (Roy et al., 2010), are insufficient from a temporal perspective in this case. Individually, neither Landsat nor MODIS data are suited for multi-scale phenological analysis in dryland areas.
An alternative to the reliance on a single inadequate dataset is to use data fusion techniques. Satellite image fusion is the synergistic blending of multiple co-located images that possess distinct yet complementary attributes; the goal is to produce a result that mitigates or transcends the individual limitations of each contributing dataset. A typical fusion combination merges low-temporal/high-spatial resolution data with high-temporal/low-spatial resolution data (e.g., Landsat and MODIS) to increase the availability of detailed imagery (Lunetta et al., 1998). Since high spatial resolution imagery is commonly used to capture spatial details on the landscape while high temporal resolution data are used to describe changes over time (Hilker et al., 2009), the fusion of the two datasets can be a powerful tool for tracking and analyzing changes of smaller-scale, dynamic elements of a landscape.
Data from the Landsat and MODIS platforms lend themselves well to data fusion studies given their similar spectral collection configurations and orbital parameters. The Landsat-7 and the MODIS Terra satellites share the same orbit, yielding imagery collected under similar viewing geometries and solar zenith and atmospheric conditions; Landsat-7 crosses the equator ~ 15 min before MODIS Terra nadir observations (Roy et al., 2008). Landsat-5 is in an identical orbit but crosses with an 8-day offset from the Terra nadir observation date. The spatial and temporal adaptive reflectance fusion model (STARFM) developed by Gao et al. (2006) takes advantage of the similarities of these disparate datasets. STARFM is an empirical fusion technique that combines Landsat 30 m data with MODIS 500 m data to provide synthetic Landsat imagery at MODIS time intervals. Unlike most other data fusion methods, STARFM produces results that are calibrated to spectral reflectance and can thus be used to track quantitative changes in surface reflectance due to phenology (Gao et al., 2006).
In this study we build upon previous work by Gao et al., 2006, Hilker et al., 2009. Our goal is to examine the efficacy of using the STARFM algorithm to generate synthetic imagery over a vegetated dryland area for the purpose of monitoring phenology changes in dryland ecosystems. We employ Landsat-5 data to evaluate the feasibility of using STARFM with a Landsat data set that is unaffected by the scan line corrector (SLC) issues of Landsat-7 ETM + data after 2003. Given the temporal mismatch of Landsat-5 and MODIS nadir acquisitions, we investigate the effect of temporal compositing and bidirectional reflectance distribution function (BRDF) adjustment on the accuracy of the STARFM imagery to determine which MODIS dataset gives the best results. We additionally examine the sensitivity of the STARFM algorithm to the timing of data pairs that establish the Landsat/MODIS spectral relationship by creating synthetic images from both early- and late-season data. We finally explore the use of STARFM for phenological purposes by comparing the peak timing and values of a vegetation index extracted from parallel STARFM, Landsat, and MODIS data time series.
Section snippets
Study site
The area of analysis is a 13,949 km2 subset of Landsat WRS-2 path 37/row 36, centered at 34°48.0′ N, 112°5.5′ W in central Arizona (Fig. 1). This area constitutes the northern extent of the Landsat scene and features broad swaths of dryland forest vegetation. It encompasses an elevation range of 814–2816 m, including both desert lowlands as well as mountain peaks west of the city of Flagstaff. The climate is semi-arid with average annual precipitation amounts ranging from 28 to 89 cm (PRISM, 2008
Results
The application of the STARFM algorithm for the Landsat/STARFM comparison yielded 60 synthetic images for the April–October 2006 time frame. A selection of the final set of synthetic images is shown in Fig. 5. The center stripe in the daily images is due to the location of the study site at the junction of MODIS overpasses. This phenomenon is clearest in the MOD09GA daily images, although variations are also visible in the 8-day composite datasets. The performances of the individual bands and
Pixel-based reflectance differences
The mean absolute difference results are comparable to those returned in the initial validation of the STARFM performance over a high-latitude coniferous forest (Gao et al., 2006). The stability of the results suggests that the underlying principles of the algorithm are sufficiently robust to accommodate the differences in imagery platform and sensors (Landsat-7 ETM + vs. Landsat-5 TM) and locations (high-latitude coniferous forest vs. dryland forest) that were introduced in this research.
Dataset differences
The
Conclusion
This research has demonstrated the feasibility of using the STARFM algorithm to create synthetic, high-resolution imagery in dryland ecosystems. The range of MODIS datasets evaluated suggests that the MODIS NBAR product is the most applicable for use with the Landsat-5 data, given the 8-day temporal offset of the Landsat-5 and Terra nadir observations. The viewing angle corrections inherent in the NBAR dataset increase the likelihood that observed spectral differences are due to phenology
Acknowledgments
We gratefully acknowledge the contributions of the two anonymous reviewers whose comments and suggestions substantially improved this paper.
References (44)
- et al.
A global overview of drought and heat-induced tree mortality reveals emerging climate change risks for forests
Forest Ecology and Management
(2010) - et al.
Land surface phenology, climatic variation, and institutional change: Analyzing agricultural land cover change in Kazakhstan
Remote Sensing of Environment
(2004) - et al.
Using a multikernel least-variance approach to retrieve and evaluate albedo from limited bidirectional measurements
Remote Sensing of Environment
(2001) - et al.
Generation of dense time series synthetic Landsat data through data blending with MODIS using a spatial and temporal adaptive reflectance fusion model
Remote Sensing of Environment
(2009) - et al.
Impacts and uncertainties of upscaling of remote-sensing data validation for a semi-arid woodland
Journal of Arid Environments
(2008) - et al.
The availability of cloud-free Landsat ETM + data over the conterminous United States and globally
Remote Sensing of Environment
(2008) - et al.
Development of an approach for generation of temporally complete daily nadir MODIS reflectance time series
Remote Sensing of Environment
(2010) - et al.
Web-enabled Landsat Data (WELD): Landsat ETM + composited mosaics of the conterminous United States
Remote Sensing of Environment
(2010) - et al.
Multi-temporal MODIS–Landsat data fusion for relative radiometric normalization, gap filling, and prediction of Landsat data
Remote Sensing of Environment
(2008) - et al.
First operational BRDF, albedo nadir reflectance products from MODIS
Remote Sensing of Environment
(2002)
Monitoring vegetation phenology using MODIS
Remote Sensing of Environment
An enhanced spatial and temporal adaptive reflectance fusion model for complex heterogeneous regions
Remote Sensing of Environment
Temperature sensitivity of drought-induced tree mortality portends increased regional die-off under global-change-type drought
Proceedings of the National Academy of Sciences of the United States of America
Desertification in central Argentina: Changes in ecosystem carbon and nitrogen from imaging spectroscopy
Ecological Applications
Ecoregions: The ecosystem geography of the oceans and continents
The Pacific Coast. In An analysis of the land base situation in the United States, 1989–2050
Regional vegetation die-off in response to global-change-type drought
Proceedings of the National Academy of Sciences of the United States of America
Reorganization of an arid ecosystem in response to recent climate change
Proceedings of the National Academy of Sciences of the United States of America
Regional climate projections
Assessing forestland area based on canopy cover in a semi-arid region: A case study
Forestry
Land surface phenology: A community white paper requested by NASA
On the blending of the Landsat and MODIS surface reflectance: Predicting daily Landsat surface reflectance
IEEE Transactions on Geoscience and Remote Sensing
Cited by (248)
A review of remote sensing image spatiotemporal fusion: Challenges, applications and recent trends
2023, Remote Sensing Applications: Society and EnvironmentImproving generalisability and transferability of machine-learning-based maize yield prediction model through domain adaptation
2023, Agricultural and Forest MeteorologyUnpaired spatio-temporal fusion of image patches (USTFIP) from cloud covered images
2023, Remote Sensing of EnvironmentROBOT: A spatiotemporal fusion model toward seamless data cube for global remote sensing applications
2023, Remote Sensing of Environment