Spatially-explicit depictions of plant production that are derived in a consistent and repeatable manner are critical to monitoring landscapes in heterogeneous arid and semi-arid grassland and savanna ecosystems (hereafter “rangelands”). Change detection techniques applied to remotely sensed data provide opportunities to identify and characterize changes in land surface conditions to assist decision-making for land management and to focus field reconnaissance efforts.
Predictions for future climate are increasing temperatures and variability in the amount and timing of rainfall in the water-limited regions of the southwestern USA [1
]. These changes in conjunction with increasing pressure on resources posed by a rapidly growing human population in this region necessitate effective, consistent, and data-driven tools to guide land management decisions and envision novel scenarios. The USA encompasses approximately 312 million hectares of rangelands, 43% of which is managed by the federal government [3
]. The millions of hectares under federal jurisdiction pose a particular challenge to meeting the need for data-driven models that are linked to ecosystem function.
Understanding vegetation dynamics is central to the assessment of rangeland resources [4
]. Over the last twenty years there has been a shift to using conceptual models that integrate non-linear vegetation dynamics. These models, known as state-and-transition models (STMs), encapsulate the notion that vegetation communities are present in multiple stable states, where the term “state” refers to the physical structure and set of ecological processes associated with a particular vegetation community. Transitions between states are catalyzed by both persistent environmental pressures such as drought or soil degradation and catastrophic disturbance [5
]. STMs are constructed according to “ecological sites” [6
]. Ecological sites are part of the Land Resource Hierarchy developed by the USDA-NRCS ( http://soils.usda.gov/survey/geography/hierarchy/
) and are distinguished according to soil type, climate and geomorphic position. Central to ecological site classification is the potential of a place to support a given vegetation community and to exhibit specific processes that lead to transitions in vegetation composition and/or structure from the expected historic state [6
State-and-transition models use various ecological site-specific terms to depict different states. We use the generalized nomenclature adopted by Steele et al.
]. Historic state
describes an historic unaltered vegetation community; an altered state
describes a vegetation community that has undergone minor changes in ecosystem structure and governing ecological processes, but the soil profile remains largely intact. A degraded
state describes a vegetation community that has undergone major changes in ecosystem structure and the ecological processes that govern transitions. Minor losses may be observed from the uppermost soil layer (A horizon). In many parts of the southwestern US, adverse changes in the vegetation structure that lead to degraded states are primarily associated with increases in woody cover and decreases in the cover of native perennial grasses to create shrub- or tree-dominated states. The most degraded state is the Bare-Annuals state, where there is widespread depletion of the A horizon and almost all vegetation has been removed except for annual pioneer species.
Establishing direct relationships between remotely-sensed data and ecological states or sites is extremely challenging in arid rangelands due to (i) the heterogeneity and sparseness of the vegetation and (ii) that fact that not all states exhibit distinct spectral signatures [8
]. We suggest that radiometric change detection techniques can serve as an effective method to evaluate land surface changes in the context of vegetation dynamics (i.e.
, state changes) predicted by STMs and to identify locations to focus field monitoring efforts. This is achieved by integrating the change surface with available spatially-explicit ancillary data.
Vegetation index differencing (VID) using moderate spatial resolution Landsat Thematic Mapper (TM) imagery is one way to achieve consistent depictions of land surface change for its multi-decadal record and ground resolved distance. VID is a radiometric method for detecting change in pixel values between dates [9
] and avoids issues associated with other change detection techniques [10
]. We sought to broadly assess land surface conditions in actively managed rangeland landscapes that occur on the interface of the Chihuahuan and Sonoran Deserts in the southwestern USA by capitalizing on the readily available Landsat 5 Thematic Mapper image archive [12
We examined changes between pre- and post-growing season Normalized Difference Vegetation Index (NDVI) [13
] values to depict the change in photosynthetic biomass over the growing season. We selected imagery for a year following persistent below-average rainfall in the region (2003) and a year following two historically high rainfall years (2009; see Figure 1
). We hypothesized that ecological states representing different degrees of degradation (i.e.
, Historic, Altered, Degraded) would respond differently to conditions of drought and abundant rainfall. We sought to answer the following questions: (1) What ecological states exhibit most positive and negative change in NDVI after periods of above- and below-average rainfall? and (2) What can we infer from the pattern of NDVI responses with respect to state-specific vegetation dynamics?
Broad-scale monitoring of land surface conditions is a pressing need in many parts of the world as the demand for multiple uses intensifies. Remote sensing change detection provides an opportunity to evaluate and identify areas of change and assist in natural resource problem solving [21
]. We used a multi-scale approach that integrating ecological state mapping using fine-resolution aerial photography with vegetation index differencing using moderate resolution satellite imagery to greatly enhance our interpretation of growing season responses to high and low rainfall. Moreover, interpretations of NDVI anomalies were informed by vegetation dynamics derived from state-and-transition models in an effort to facilitate ecosystem inventory and monitoring efforts.
While we can make inferences regarding NDVI changes using knowledge of vegetation dynamics in this region, this method must be developed further and correlated with field observations to verify our inferences. A traditional accuracy assessment was beyond the scope of this initial study. This is primarily because we lack ground data suitable for determining the degree of change in vegetation indicated by VID over such an expansive area of which much is difficult to access. Yet, future efforts are planned that will incorporate field visits to areas that demonstrated dynamic responses to drought and periods of high rainfall. The effort will incorporate on-the-ground data collection with on-going analysis of fine spatial resolution imagery image analysis to refine and inform the ecological state mapping process. The classification of ecological sites using fine spatial resolution imagery is a developing field that is most successfully accomplished with iterative refinement and development as field data are compiled [8
]. We uphold the integrity of the ecological state maps generated in this study with expert knowledge and experience in interpreting vegetation patterns and associations with geomorphology discernible on the digital ortho-quarter quadrangle imagery; however, we acknowledge that the mapping of ecological state polygons represents a source of uncertainty. Where this expert knowledge is not available, VID interpretations could still be informed by readily-available ecological site polygons.
We present a case study to demonstrate how a multi-scale approach to change detection can effectively focus field reconnaissance efforts that in itself provides insight where field data are presently lacking. Further development can be greatly enhanced by incorporating longer time series such as Landsat data available through the Web-enabled Landsat Data (WELD) project [23
We contend that even in the absence of field data or expert knowledge, the use of VID is valuable for prioritizing sites for field visits. Vegetation communities occurring in altered states have been shown to respond well to management intervention whereas a vegetation community in a highly degraded state is often beyond economical means of intervention. We recommend that locations where the vegetation communities are in historic or altered states and which appear as growing season NDVI anomalies (positive or negative) should be prioritized for field visits over locations showing less significant changes in NDVI. In this manner, spatially-explicit depictions of areas with potential for vegetation/ecological state transition would greatly enhance the effectiveness of field and management efforts across millions of acres of federal lands.
All remote sensing protocols designed to provide data needed for decision-making have strengths, weaknesses, and situations for which they function optimally. That the 2009 growing season NDVI VID values were not normally distributed does not compromise our ability to identify areas in the tails of the distribution to identify patterns and guide field efforts for broad-scale landscape monitoring. When research objectives or management needs require multiple depictions of land surface condition (either multiple dates or multiple sites at one time), growing season NDVI values must be standardized. If emphasis is placed on a mapping effort for a site at one point in time, the user could alternatively rank the VID values and choose those at the upper and lower ends of the distribution. This modification and/or non-parametric techniques could be used to avoid violating assumptions associated with normally distributed data.
Coppin et al.
] and Singh [11
] provide an effective treatment of logistical considerations as well as advantages and disadvantages of different change detection methods. The use of imagery collected at different spatial resolutions and the combination of manual (ecological state mapping) and automated (NDVI difference images) remote sensing methods provided an intuitive data product. This application of vegetation index differencing (VID) combines two benefits noted by Coppin et al.
]; first, vegetation indices are more closely related to land surface changes than individual image bands, and second, VID is capable of detecting both abrupt and progressive changes in the land surface. The latter point is of tremendous benefit to those seeking to identify indicators that portend ecological state transitions. In addition, the direct use of radiometric data (i.e.
, spectral vegetation indices) and translation to growing season change in NDVI circumvents image classification errors associated with multi-date comparisons, or other post-processing and did not rely on robust relationships between NDVI and biophysical parameters, of which a notable lack exists in drylands. Lastly, there is flexibility in this approach in both the selection of the appropriate Z-score threshold to define “high” and “low” responses and the scale of observation, i.e.
, minimum patch size. Both refinements for thresholds and patch size should reflect the specificity and focal scale required to achieve the management or research objective [21
]. The ability to modify these factors in a decision-making framework is highly valued by land managers [25
There are research applications for which remotely sensed imagery assist, but do not fulfill decision-making needs and requirements [21
] and the strengths and limitations should be duly noted. Land managers and decision-makers seek remote sensing tools that provide products relevant to and consistent with STM concepts [6
]. This is a compelling challenge from two perspectives. The remote sensing community is needed to augment the knowledge regarding the accuracy and suitability of the full suite of change detection algorithms not presented here, e.g., [11
] to promote understanding of which techniques are best suited for different research applications. Land managers and their technical collaborators are challenged to identify existing indicators or modifications thereof that are commensurate with products derived from remotely sensed data [25
]. Only with contributions from both communities and effective dialogue between them will the full potential of remote sensing for natural resource management decision-making be realized.