Land-Cover Change Detection Using Multi-Temporal MODIS NDVI Data

Monitoring the locations and distributions of land-cover change is important for establishing links between policy decisions, regulatory actions and resulting land-use activities. Past activities including two-date change detection efforts using Landsat data have tended to be performance limited for applications in biologically complex systems. This study explored the use of 250 m multi-temporal MODIS NDVI 16-day composite data to provide an automated change detection and alarm capability on a one year time-step for the Albemarle-Pamlico Estuary System (APES) region of the US. Detection accuracy was assessed for 2002 at 88%, with a reasonable balance between change commission errors (21.9%) and change omission errors (27.5%) and Kappa coefficient of 0.67. Annual change detection rates across the APES over the entire study period (2002-2005) were estimated at 0.7% per annum and varied from 0.4% (2003) to 0.9% (2004). Regional variations were also readily apparent ranging from 1.6% to 0.1% per annum for the tidal water and mountain ecological zones, respectfully. This research included the application of an automated protocol to first filter the MODIS NDVI data to remove poor (corrupted) data values and then estimate the missing data values using a discrete Fourier transformation technique to provide high quality uninterrupted data to support the change detection analysis. The methods and results detailed in this article apply only to non-agricultural areas. Additional limitations attributed to the coarse resolution of the NDVI data included the overestimation of change area that necessitated the application of a change area correction factor. Page 2 of 29 INTRODUCTION Land-cover (LC) composition and change are important factors that affect ecosystem condition and function. These data are frequently used to generate landscape-based metrics and to assess landscape condition and monitor status and trends over a specified time interval (Jones et al., 1997). The use of satellite-based remote sensor data has been widely applied to provide a cost-effective means to develop LC coverages over large geographic regions. Past and ongoing efforts for generating LC data for the United States have been implemented using an interagency consortium to share the substantial costs associated with satellite data acquisition, processing and analysis. The first fine resolution National Land-Cover Data (NLCD) set was developed for the conterminous United States using Landsat Thematic Mapper (TM) imagery collected between1991-1992 (Vogelmann et al., 1998). Currently, the NLCD-2001 is under development for all 50 States and the Commonwealth of Puerto Rico (Homer et al., 2004). Although the NLCD-2001 is expected to provide the most current national LC database currently available, the required development time will result in approximately a five year delay between data collection and product availability (Homer et al., 2004). Ideally, a more current NLCD product would be available to support on-going environmental assessment and policy decisions. To achieve this goal in a cost-effective manner, one possible approach would be to identify areas of LC change occurring subsequent to 2001 and append the NLCD-2001 for only those areas that have undergone change. The updated NLCD would not only provide the user with current LC data, but could also be used to identify both the pattern and nature of changes that had occurred between dates of interest. Other advantages associated with the updated editing change images would include, (a) potentially substantial cost savings, (b) no introduction Page 3 of 29 of new classification errors for non-change areas, and (c) minimization of registration errors that typically limit the overlay of multiple date coverages (post-classification) to support change detection analysis. Major impediments that need to be overcome for the development of practical methods for large-area change detection monitoring using remote sensor data include both the elimination or significant reduction of interand intra-annual vegetation phenology mediated errors and robust methods for error assessment. There is also an expanding need for continuous data streams to support the development of the next-generation of spatially distributed landscape process models that would incorporate both greater spatial resolution and higher frequency simulations (time-steps). Specifically, remote sensing can provide estimates of standing Photosynthetically Active Biomass (PAB) that can serve as a nearly continuous measurement surrogate of primary productivity distributions and provide accurate measurement metrics of the timing and duration of important physiological events (i.e., green-up, duration of growing season, and senescence). Desirable diagnostic data include numerous Vegetation Indices (VIs) such as the Normalized Difference Vegetation Index (NDVI), fraction of Photosynthetically Active Radiation (fPAR), and Leaf Area Index (LAI). To maximize both performance and utility, future LC change detection methods that incorporate PAB data streams could provide a robust approach for the near real-time monitoring of LC change events while simultaneously supporting the development of landscape indicators, modeling of important to landscape processes, and the forecasting of future LC change distributions. Page 4 of 29 BACKGROUND Past spectral-based change detection techniques have tended to be performance limited in biologically complex ecosystems due, in larger part, to phenology induced errors (Lunetta et al., 2002). An important consideration for LC change detection is the nominal temporal frequency of remote sensor data acquisitions required to adequately characterize change events (Lunetta et al., 2004). Ecosystem-specific regeneration rates are an important consideration for determining the required frequency of data collections to minimize omission and commission errors. As part of the natural processes associated with vegetation dynamics, plants undergo intra-annual cycles (phenology). During different stages of vegetation growth, plant structures and associated pigment assemblages can vary significantly. Our ability to identify vegetation classes using remote sensor systems is a result of wavelength specific foliar reflectance (0.76-0.90 μm), pigment absorptions (0.45-0.69 μm), and foliar moisture content (1.55-1.75 μm). The same vegetation type can appear significantly different and different types similar, at various stages during intra-annual growth cycles. Also, significant difficulties in evaluating the performance of change detection methods result from the inability to adequately characterize outcome accuracies (Khorram et al., 1999). Unlike typical LC classification assessment that requires only single date validation data, change detection validation data must be available for multiple dates to provide sufficient change event documentation (i.e., before and after). In particular, the characterization of change omission errors represent an especially difficult challenge (Lunetta et al., 2002 and 2004). Because LC change (conversion) is a relatively rare occurrence over a large unit area it is problematic to derive a robust estimate of change omission errors. Consequently, standard LC accuracy Page 5 of 29 assessment methods tend to over estimate the actual performance. Rather than relying on the traditional percent accuracy statistic, a more useful representation of performance is the Kappa coefficient (Cohen, 1960) and characterization of the percentage of both change and no-change omission and commission errors (Lunetta et al., 2002). Remote sensing change detection techniques can be broadly classified as either preor post-classification change methods. Pre-classification methods can further be characterized as being spectral or phenology based. Originally, the post-classification approach was considered to be the most reliable approach and was used to evaluate emerging methods (Weismiller et al., 1977). Factors that limit the application of post-classification change detection techniques can include: cost, consistency, and error propagation (Singh, 1989). Numerous pre-classification change detection approaches have been developed and refined to provide optimal performance over the greatest possible range of ecosystem conditions. These semi-automated digital data processing approaches include image-based composite analysis (Weismiller et al., 1977), and principal components analysis (PCA) (Lillesand and Keifer, 1972; Byrne et al., 1980; Richards, 1984). The most commonly applied data transformations applied include band ratioing, NDVI, and the tasseled-cap transformation (Crist, 1985; Jensen, 2005). Recent techniques have been applied that can interpret data transformation results using change vector analysis (CVA) to indicate the magnitude and nature of change (Lambin and Strahler, 1994). Research evaluating the comparative performance of various LC change detection methods has indicated that no uniform combination of data types and methods can be applied with equal success across different ecosystems (Lu et al., 2004). Cohen and Fiorella (1998) found that composite analysis out performed both image differencing and CVA in a two-date Page 6 of 29 experiment for detecting conifer forest change using Landsat Thematic Mapper (TM) imagery in the Oregon Cascades. Lyon et al. (1998) reported that the NDVI was the best performing Vegetation Index (VI) for detecting LC changes in the biologically complex vegetation communities in Chiapas, Mexico. Subsequently, Lunetta et al. (2002) determined that two-date NDVI differencing and Multiband Image Differencing (MID) both performed poorly in a biologically complex vegetation community in North Carolina. Analysis of LC data for the country of Mexico using three dates of Landsat Multi-Spectral Scanner (MSS) imagery over a two-decade (1972-1992) study period demonstrated that ecosystem variability combined with classification errors precluded scene-wise or pixel-wise change detection (Lunetta et al., 2002). Past studies have demonstrated the potential of using NDVI to study vegetation dynamics (Townshend and Justice, 1986; Verhoef et al., 1996), illustrate the value of using high temporal resolution MSS imagery to monitor changes in wetland vegetation (Elvidge et al., 1998) and document the importance of image temporal frequency for accurately detecting forest changes in the southeastern United States ( Lunetta et al., 2004) . Time series data analysis using remote sensor data have largely focused on the use of coarse resolution ($1.0 km) AVHRR data to document land-cover and analyze vegetation phenology and dynamics (Justice et al., 1985; Townshend and Justice, 1986; Justice et al., 1991; Loveland et al., 1991). With the advent of MODIS NDVI 250 m data, time series data analysis can be adapted for higher (moderate) resolution applications. However, the utility of the MODIS NDVI data products are limited by the availability of high quality (e.g., cloud free) data Jin and Sader (2005). The availability of high quality data is a critical factor in determining application utility. To best deal with the data quality issues researchers have incorporated a number of processing techniques including Page 7 of 29 weighted regression smoothing (Li and Kafatos, 2000), Fourier and wavelet transformation filtering (Sakamoto et al., 2005), weighted least squares (Reed, 2006) and wavelet feature extraction (Bruce and Mathur, 2006). Study Objectives The purpose of the research described here was to investigate the feasibility of using Moderate Resolution Image Spectroradiometer (MODIS) derived NDVI data to identify change areas on an annual time-step. The goal was to develop an automated change detection alarm capability for regional scale applications using a nearly continuous high quality data stream that could also be used to support phenology-based cover type classification and further the study agricultural land-use dynamics. This research effort focused only on non-agriculture cover types. Because of the unique issues associated with differentiating between agricultural land-use and land-cover conversions, methods for agricultural conversions to non-agricultural cover types (e.g., urban and forest) are being developed separately from the effort described in this article. Study Area This research was conducted across the 52,000 km Albemarle-Pamlico Estuary System (APES) located in North Carolina and Virginia (Figure 1). The study area includes diverse ecoregion types ranging from tidal plain (east) to the Blue Ridge Mountains (west). Landscape “patch” sizes are relatively small, with approximately 95% of the study area exhibiting heterogeneous cover within an individual MODIS NDVI (250 m) pixel. Additionally, biological diversity and vegetation regrowth rates are at the high end of the spectrum for locations within Page 8 of 29 the conterminous United States (Currie and Paquin, 1987). The composition of the major upland-cover types throughout the area as of 2001 were approximately 51% forest, 23% agriculture, 14% water, 7% urban and barren, 4% grassland, and 1% herbaceous wetland (NASS, 2002; Homer et al., 2004). The predominate LC conversions that occur in the APES are associated with deforestation for both urban development and forest harvesting (clear cutting). METHODS Methods incorporated in this study included the application of an automated MODIS NDVI time series and reference databases for the APES study area to support multi-temporal imagery analysis and accuracy assessment. MODIS NDVI data preprocessing was conducted to provide a filtered (anomalous data removed) and cleaned (excluded data values estimated) uninterrupted data stream to support multi-temporal (phenological) analysis. Figure 2 illustrates the complete data pre-processing flow that is detailed below. Total annual NDVI values for each 250 m grid cell within the study area (20012005) were compared on an annual basis to identify those cells exhibiting greater than specified threshold values (decrease in NDVI) and were labeled as LC conversion areas. An accuracy assessment was conducted using historical aerial photographs to document the occurrence or non-occurrence of conversion events to generate an overall accuracy and quantify both commission and omission errors. MODIS NDVI 16-day composite grid data (MOD13Q1) in HDF format were acquired between February 2000 and December 2005 from the NASA Earth Observing System (EOS) data gateway. Details documenting the MODIS NDVI compositing process and Quality Assessment Science Data Sets (QASDS) can be found at NASA’s MODIS web site (MODIS, Page 9 of 29 1999). NDVI data were subset to the APES study boundary (10 km buffer), re-projected from a sinusoidal to an Albers Equal-Area Conic projection, using a nearest-neighbor resampling routine, and entered into a 250 m x 250 m grid cell multilayer image stack. Separate data stacks were developed for both the original NDVI data and QASDS. The NDVI data stack was first filtered to eliminate anomalous high (hikes) and low (drops) values and then filtered for a second time using the QASDS ratings to remove poor quality data values from the NDVI data stack. Hikes and drops were effectively eliminated by removing data values that suddenly decreased or increased and then immediately returned to near the previous NDVI value. The threshold for the removal of pseudo hikes and drops was set at ±0.15% to achieve the best setting (determined qualitatively) to eliminate most all anomalous points, while not inadvertently removing good data points, thus producing a smoother temporal profile. The MODIS QASDS data quality ratings were then applied to retain only those pixels rated as “acceptable” or higher. The filtered data were then transformed into frequency domain using a discrete Fourier transformation and the signal and noise spectrum separated (Roberts et al., 1987; Azzali and Menenti, 2000; Roerink and Menenti, 2000). The removed (corrupted) NDVI data points were estimated from the frequency domain signal spectrum using a nonlinear deconvolution approach described by Roberts et al. (1987) to estimate complete “filtered and cleaned” NDVI temporal profiles for each pixel within the APES (Figure 3). The variable nature of water bodies (ie., turbidity and water level fluctuations) tended to confound change analyses and thus were excluded to reduce commission errors (false positives). A water mask developed using Landsat Enhanced Thematic Mapper plus (ETM+) images (year 2000) was used to delineate open water areas for exclusion from further change analysis. ETM+ Page 10 of 29 pixels with negative NDVI values were identified as open water and an overlay of water pixels (30 m) was applied to the 250 m APES grid cells was performed and the percent open water area within each cell was calculated for the entire APES. Grid cells containing $30% open water were included in the APES water mask and those containing between <30% open water were used to create the APES transitional water mask. Approximately 14 % of the APES study area was included in the open water mask and 5.1% in the transitional water mask, leaving 80.8% (approximately 42,016 km) of the APES defined as upland (non-water). Intensive land-use activities associated with agricultural crop rotations often confound LC conversion determinations and can frequently result in unacceptable levels of false positives (change commission errors). Accordingly, an initial stratification step was implemented to subset the agricultural versus non-agricultural pixels for independent analysis to effectively minimize the potential for omission change errors in non-agricultural areas and commission change errors for agricultural areas. The identification of agricultural cells was accomplished by manually interpreting agricultural cover using the same ETM+ data used to create the above described water mask. All APES 250m grid cells containing TM20% agricultural cover were included in the agricultural subset for the upland areas (Knight et al., 2006). Change analysis for the agricultural areas will be performed independent of this effort for agriculture cells throughout the APES study area. Total annual NDVI values were computed for all upland grid cells using the filtered and cleaned data stack. NDVI difference values for non-agricultural areas were calculated corresponding to each individual upland cell for all consecutive study years (i.e., 2001-2002, 2002-2003, 2003-2004, 2004-2005). The total NDVI difference output data for the four change Page 11 of 29 periods beginning with 2001-2002 exhibited an approximately normal distribution about the mean (μ=0), which, for this application, represented no change in NDVI between T1 and T2. Standard normal distribution statistical analysis was performed to identify those cells that had the greatest reduction in cumulative NDVI for each change period. Four negative (decrease) change thresholds were selected corresponding to a range of z-value probabilities (i.e., 2.0, 2.5, 3.0, 3.5). This range of values were selected because they produced appropriate estimates of annual change values based on previous change rate studies for the area (Loveland, et al., 2002; Lunetta et al., 2004). Accuracy Assessment Reference data to support the accuracy assessment consisted of available historic aerial photography over a two county area (Durham and Wake) in the Piedmont region of the study area. The accuracy assessment was limited to these two counties because they were the only counties within the APES with sufficient historical aerial photography to support an assessment. Contained within these two counties are the metropolitan areas of Raleigh and Durham. All major cover types found throughout the APES study area are represented within the two counties. For each validation point, a coordinate based search was performed to identify all available historical photographs. Those sampling points with available photographs that both predated and postdated the change or no-change event were identified for further processing. Using the two dates of photographs and NDVI temporal profile the interpreter determined the usability of the photographs as a reference source for the assessment. If the photographs spanned the pre and post time of change event indicated as indicated by the temporal profile then Page 12 of 29 the photographs was deemed appropriate for assessment purposes. However, if the timing of the change indicated by the temporal profile was not within the dates of photographs collection the validation point was eliminated. For each assessment point with acceptable photographs, interpretations were conducted by first delineating the cell boundary on the digital images for both dates. The cover type(s) and percent coverage within the cell were interpreted for both dates and recorded. Interpretations were conducted by two interpreters. As part of the quality control (QC) process the lead interpreter reviewed all interpretations for consistency and data recording completeness. Also, for quality assurance (QA) purposes all reference image pairs were electronically filed to support subsequent analyses (Figure 4). Additionally, for a subset of the change points (n=52) and nochange points (n=44) sufficient image data was available to document with a high level of certainty the percentage of change that occurred within the cells (Figures 6 and 7). Reference data were developed using a stratified random sampling approach that included both the change and no-change areas. Change areas were weighted with a higher proportion of samples because the vast majority of the study area was no-change for the reference study period. A validation was performed only for change that occurred during 2002 due to the lack of available photographs for later time periods. A total of 1,137 samples were successfully validated corresponding to change (n=318) and no-change (n=819) pixels. Note that the no-change pixels were over-sampled to provide a sufficient sample size to assess nochange omission errors (Table 1). Page 13 of 29 RESULTS NDVI Profiles A qualitative analysis of the temporal profiles corresponding to phenological endmembers was performed to provide an initial performance assessment of the automated MODIS NDVI protocol described above for producing high quality time series data to support subsequent change detection analysis. Figure 5 illustrates the four principal endmembers for the Piedmont region of the study area. The patterns and time sequencing of phenological events for both the deciduous and coniferous forest types demonstrate the robust nature of the profiles derived from the automated data processing procedures. Also, the substantially lower NDVI for the impervious compared to chlorophyll-bearing endmembers is what would be expected, and the sawtooth pattern provided some indication of the inherent noise level in the NDVI product. The agricultural endmember serves to demonstrated the conflict between agricultural land-use activities and land-cover conversions. For example, over the first two year period (2000 and 2001) it appeared that the same (or similar) crop type was grown that exhibited a multiple maturation growth pattern. During year three (2002) the field appeared to be planted later in the season with an alternative crop type, then was apparently replanted in 2003 and 2004 with a new crop type possibly as part of a crop rotation cycle. This change in phonological cycle demonstrated the potential for land-cover change false positives in agricultural areas. A unique approach to supplement the validation process is illustrated in Figure 4 which shows an example of aerial photography in combination with the five year NDVI temporal profile for an accuracy assessment sampling site (reference pixel) before and after a conversion event. Much of the forest that was present in the May 1, 2000 image had been subsequently Page 14 of 29 cleared by the date of the companion image on September 1, 2002. By visually interpreting the NDVI temporal profile, it is apparent that the forest clearing began approximately mid-August 2001 and was completed by early December 2001. The area was in a transitional state throughout calendar year 2002 and appears to have reached a new steady state beginning in 2003 and throughout 2004. With the subsequent addition of the calendar year 2005 temporal profile the achievement of a new steady state condition should be confirmed. (Subsequent 2005 field reconnaissance has revealed that the site was converted to a condominium complex.) Change Detection The change detection results for non-agricultural areas are presented in Table 2. A standard deviation (SD) threshold (TH) of 2.5 provided the best accuracy at 87.5% and a Kappa coefficient of 0.67. This result was statistically significant compared to the TH factors of 2.0 (p=0.05) and 3.5 (p=0.27); however, it was not significantly different from 3.0 (p=0.72). Most importantly, the TH of 2.5 provided the best balance between change commission errors (21.9%) and omission errors (27.5%). Alternatively, the TH of 3.0 could be applied if the goal were to bias for lower change commission errors (19.4%) at the expense of higher omission errors (34.2%). Estimates of the area extent of change for the APES varied by a factor of 3.3x across the entire TH factor range; however, the TH 2.5 to 3.0 range varies only by 0.3x (Table 3). An important consideration for the application of the change detection procedures presented here is the sub-pixel sensitivity or the minimal detection limit associated with the approach. The sub-pixel sensitivity is also important to accurately determine the area extent of land-cover conversions. For example, because the majority of identified changes are sub-pixel Page 15 of 29 in extent, a correction factor must be applied to capture a true area estimate of change. Figure 6 illustrates the distribution of sub-pixel sensitivity for Wake County, NC. Approximately 96% of the identified change was at the sub-pixel level, averaging 40.4%. To best estimate the minimum detection limit, the distribution of percent change associated with change omission pixels can be used (Figure 7). The distribution shows that approximately 71% of the change omission pixels had #20% change. This equates to an effective minimum detection limit of approximately 1.5 ha. Although pixels with #20% change are detected, the detection accuracy drops substantially (increased omission errors) below 1.5 ha (Figure 6). The impact of using non-adjusted change detection area extent values is clearly apparent between the raw versus adjusted values (Table 3). Comparing the adjusted values across the study area reveals a large (5.0x) variation in land-cover change by APES major ecological regions (Table 4). The tidal ecological region had the greatest adjusted average annual change over the four year change period (1.1%), followed by the coastal plain (0.8%), Piedmont (0.5%) and mountain (0.2%). The change rate for the Piedmont (0.5%); however, the adjusted average change rate in the Raleigh metropolitan area (a subset of the Piedmont) was substantially higher at 0.9%. These results indicate that the tidal and coastal ecological regions, along with major metropolitan areas, would probably require more frequent change updates than the mountain region to provide relevant change patterns and current land-cover data sets. Additionally, the tidal ecoregion had undergone the greatest change during 2004 (1.6%). An additional analysis was conducted to better understand the relationship between cover composition, change distributions, and detection accuracies (Table 5). Although coniferous forest cover represented only 14% of the APES, 44% of all changes were associated with this Page 16 of 29 cover type and were a result of the intensive silviculture operations located primarily in the tidal and coastal ecoregions. This change rate was approximately 3x that of the deciduous and mixed forest and urban and barren areas. Accuracies were consistently in the low 90s for all reported cover types with the exception of the urban and associated barren cover types which were substantially lower at 83%. A high resolution graphic of the Raleigh metropolitan area was developed to further evaluate both the utility and performance of the annual change products (Figure 8). The graphic uses Landsat ETM+ panchromatic images (15 m) as the backdrop to aid in the interpretation of the MODIS NDVI derived annual change product. Qualitatively, the change patterns appear to match very well with known development patterns over the three year period (2002-2005). Individual residential communities, shopping centers and major road construction projects are discerned corresponding to expected time period (year). Also, the variability in change area for the study period was consistent with known economic development patterns. The drop in change area for 2003 to 0.4% (60% decrease) corresponded to the economic downturn in the technology sector; which, resulted in a dramatic slowdown in new commercial construction permits (250% dollar value reduction) in 2002 (Wake County, 2005). Commercial construction permits began to rebound in 2003 and residential construction permits increased dramatically (35% increase) in 2004, resulting in a rebound back to a 1.0% change in 2004 (Table 4). It should be noted that change is graphically over-represented (Figure 8) by approximately 2.5x, because the average pixel change area was only 40.4% (Figure 6). Page 17 of 29 DISCUSSION Articulating a working definition of what constitutes LC change is an important first step for the implementation of change detection studies and applications involving the use of remote sensor data. In this study, we made the distinction between agricultural land-use activities and land-cover change to first provide a hierarchical segmentation of our change detection procedures. Here the segmentation has been applied to reduce land-use mediated false positives (change commission errors). Additionally, this segmentation approach will be applied to support the future development of specific analytical methods for monitoring agricultural land conversions to non-agricultural cover types and to determine the potential for providing useful information such as crop type and the monitoring of rotational cycles. The implementation of segregated agricultural and non-agricultural analytical protocols should become a practical approach for the automating change detection for many locations within the US that exhibit similar properties to the APES. Additionally, this approach may be optimized when applied in a hierarchical approach in combination with the ongoing development of agricultural maps by the US Department of Agriculture (USDA) National Agriculture Statistical Service (NASS) based on Landsat ETM+ imagery (NASS, 2002.). The importance of quantifying change omission errors is critical to the development of a robust assessment of change detection performance. Absent sufficient data documenting change omission errors, the tendency would be to bias the analysis to reduce change commission errors, while potentially increasing omission errors to unacceptable levels (Table 2). Because landcover conversions typically represent only a small percentage of large geographic regions (e.g., 0.5% 1.0%), the sampling of no-change areas at sufficient intensity to generate estimates of Page 18 of 29 change commission errors represents a significant problem. This sampling problem is twofold, first a sample drawn from a no-change classified area has a very low probability of being a change omission pixel; and second, the probability of finding adequate reference data to determine the status of a sample site (i.e., change or no-change) is very low. In this study we selected approximately 2.6x more samples from the no-change bin to support the determination of change omission errors. With the anticipated release of the NLCD-2001 data scheduled for calendar year 2006, efforts can be initiated to create annual change coverages beginning with 2002 using MODIS NDVI 250 m data for numerous regions of the US where this approach may be appropriate, providing supplemental information to support the interpretation and application of NLCD-2001 products. At the most basic level, a MODIS NDVI-derived annual change detection alarm product could be applied to identify those areas that have likely undergone land-cover change subsequent to 2001. Also, it may be possible to determine the likely outcome of the conversion, subsequent to the return to a new steady state based on the application of MODIS NDVI temporal profile match filtering techniques currently under development (Knight et al., 2006). An additional advantage of the MODIS NDVI change detection procedures developed in this study is the capability for processing filtered and cleaned NDVI temporal profiles that track vegetation phenology on a nearly continuous basis to support the development of regional scale landscape process models. For non remote sensing scientists, the NDVI temporal profiles provide a particularly insightful data presentation that can readily be interpreted without any formal remote sensing training. Also, they provide a data presentation that can easily be exploited to extract data values of potential interest using widely available desktop software Page 19 of 29 tools (i.e., phenological metrics). The potential ease of sharing these multi-temporal NDVI products with the general scientific community may hasten the future integration of these products by other disciplines and potentially lead to additional unanticipated applications. CONCLUSIONS The availability of no-cost MODIS NDVI data and new automated data processing techniques that provide high quality continuous time series data, represent a major advancement for the automated monitoring annual land-cover change and vegetation condition over large geographic regions. Major advantages of the NDVI-based change detection approach presented here include (a) robust results, (b) nominal computational requirements, (c) automated data processing protocols, (d) annual change alarm product capability, and (e) rapid product delivery. Additional advantages include the portability of intermediate data products and results to the non remote sensing scientific community. This approach is particularly attractive due to the availability of no cost MODIS data and the very low cost associated with data processing. These advantages are in sharp contrast to the traditional Landsat data based approaches that are comparatively data and computationally expensive. The increased temporal resolution of the MODIS NDVI 250 m data has a significant advantage over traditional Landsat data for both capturing the actual timing of the change event and the subsequent monitoring of the recovery to the next steady state. Disadvantages and limitations of the MODIS NDVI based change detection approach are substantially associated with the moderate spatial data resolution (250 m). In particular, change events less than approximately 1.5 ha will have a low probability of being detected. One impact Page 20 of 29 of this resolution limitation is potentially poorer accuracies for urban areas that tend to change at finer scales (Table 5). Also, because of the sub-pixel sensitivity of the technique, there is an inherent and significant overestimation (approximately 2.5x) of change area extent that must be corrected to provide reasonable change rate estimates. This issue is particularly problematic for the graphic representation of change extent, because there is no practical way to directly compensate for the overestimations. The spatial resolution of the MODIS NDVI data significantly limits their use for certain applications such as the monitoring of change in riparian buffers zones and urban areas and the monitoring of other relatively fine scale conversion events that may be associated with high value ecological resources. Additionally, the method described in this article will not provide robust results for areas in intensively managed landscapes (e.g., agricultural land). ACKNOWLEDGMENTS The authors would like to acknowledge the photographic interpretation support provided by Jeffery Swain and Mark Murphy in support of the accuracy assessment effort described in this article. The U.S. Environmental Protection Agency funded and conducted the research described in this paper. It has been subject to the Agency’s programmatic review and has beenapproved for publication. Mention of any trade names or commercial products does notconstitute endorsement or recommendation for use. Page 21 of 29REFERENCESAssali, S., and Menenti, M. (2000). Mapping vegetation-soil-climate complexes in southern Africa using temporal Fourier analysis of NOAA-AVHRR NDVI data. Int. J. Remote Sensing, 21(5), 973-996. Bruce, L.M., Mathur, A., and Byrd, Jr., J.D. (2006). Denoising and wavelet-based feature extraction of MODIS multi-temporal vegetation signatures. GIScience & RemoteSensing, 43(1), 67-77. Byrne, G.F., Crapper, P.F., & Mayo, K.K. (1980). Monitoring land-cover change by principalcomponent analysis of multitemporal Landsat data, Remote Sens. Environ., 10, 175-184. Crist, E.P. (1985). A TM tasseled cap equivalent transformation for reflectance factor data,Remote Sens. Environ., 17, 301-306. Cohen, J. (1960). A coefficient of agreement for nominal scales. Educ. Psycol. Measurement,20(1), 37-46. Cohen, W.B., & Fiorella, M. (1998). Comparison of methods for detecting conifer forest changewith Thematic Mapper imagery. In: Remote Sensing Change Detection: EnvironmentalMonitoring Methods and Applications (R.S. Lunetta and C.D. Elvidge, Editors), AnnArbor Press, Chelsea, MI, 318 p. + Illustrations, and co-published in Europe by Taylor Page 22 of 29and Francis, UK. Currie, D.J., and Paquin, V. (1987). Large-scale biographical patterns of species richness of trees. Nature, 329, 326-327. Elvidge, C. D., Miura, T., Jansen, W.T., Groeneveld, D.P., & Ray, J. (1998). Monitoring trends in wetland vegetation using a Landsat MSS time series. In: Remote Sensing ChangeDetection: Environmental Monitoring Methods and Applications (R.S. Lunetta and C.D.Elvidge, Editors), Ann Arbor Press, Chelsea, MI, 318 p. + Illustrations, and co-publishedin Europe by Taylor and Francis, UK. Homer, C., Huang, C., Limin, Y., Wylie, B. & Coan, M. (2004). Development of a 2001National Land Cover Database for the United States. Photogrammetric Engineering &Remote Sensing, 70(7), 829-840. Jensen, J.R. (2005). Introductory Digital Image Processing: A Remote Sensing Perspective 3Edition. Upper Saddle River, NY, Prentice Hall, 525 p. Jin, S., and Sader., S.A. (2005). MODIS time-series imagery for forest disturbance andquantification of patch effects. Remote Sens. Environ., 99, 462-470. Jones, B.A., Ritters, K.H., Wickham, J.D., Tankersley Jr., R.D., O’Neill, R.V., Chaloud, D.J., Page 23 of 29Smith, E.R., & Neale, A.C. (1997). An Ecological Assessment of the United States MidAtlantic Region: A Landscape Atlas, U.S. Environmental protection Agency, Report No. EPA/600/R-97/130, U.S. Printing Office, Washington, DC, 104 p. Justice, C.O., Townshend, J.R.G., Holben, B.N., and Tucker, C.J. (1985). Analysis of the phenology of global vegetation using meteorological satellite data. Int. J. Remote Sensing, 6(8) 1271-1318. Justice, C.O., Townshend, J.R.G., Kalb, V.L. (1991). Representation of vegetation bycontinental data sets derived from NOAA-AVHRR data. Int. J. Remote Sensing, 12(5)999-1021. Khorram, S.K., Biging, G.S., Chrisman, N.R., Colby, D.R., Congalton, R.G., Dobson, J.E.,Ferguson, R.F., Goodchild, M.F., Jensen, J.R., & Mace, T.H. (1999). AccuracyAssessment of Remote Sensing-Derived Change Detection, American Society forPhotogrammetry and Remote Sensing, Monograph Series, ISBN 1-57083-058-4, 64 pp. Knight, J.K., Lunetta, R.L., Ediriwickrema, J., and Khorram, S. (2006). Regional Scale Land-Cover Characterization using MODIS-NDVI 250 m Multi-Temporal Imagery: A Phenology Based Approach. GIScience and Remote Sensing (Special Issue on MultiTemporal Imagery Analysis), 43(1), 1-23. Page 24 of 29Lambin, E.F., & Strahler, A.H. (1994). Change-vector analysis in multispectral space: A tool to detect and categorize land-cover change processes using high temporal-resolution satellite data. Remote Sens. Environ., 48, 231-244. Li, Z., and Kafatos, M. (2000). Interannual varibality of vegetation in United States and its relation to EL Nino/southern oscillation. Remote Sens. Environ., 71, 239-247. Lillesand, T.M., and Keifer, R.W. (1972). Remote Sensing and Image Interpretation. JohnWiley and Sons, Second Edition, 721 pp. Loveland, T.R., Merchant, J.W., Ohlen, D.O., & Brown, J.F. (1991). Development of a land-cover characteristics data base for the conterminous U.S. Photogrammetric Engineering& Remote Sensing, 57(11), 1453-1463. Loveland, T.R., Sohl, T.L., Stehman, S.V., Gallant, A.L., Sayler, K.L., and Napton, D.E. (2002).A strategy for estimating the rates of recent United States land-cover changes.Photogrammetric Engineering & Remote Sensing, 68(10), 1091-1099. Lu, D., Mausel, P., Brondizio, E., and Moran, E. (2004). Change detection techniques. Int. J.Remote Sensing, 20(12), 2365-2407. Lunetta, R.S., Alvarez, R., Edmonds, C.M., Lyon, J.G., Elvidge, C.D., Bonifaz, R., García, C., Page 25 of 29Gómez, G., Castro, R., Bernal, A., & Cabrera, A.L. (2002). An assessment of NALC/Mexico land-cover mapping results: implications for assessing landscape change. Int. J. Remote Sensing, 23(16), 3129-3148. Lunetta, R.S., Congalton, R.G., Fenstermaker, L.K., Jensen, J.R., McGwire, K.C., & Tinney, L.R. (1991). Remote sensing and geographic information system data integration: error sources and research issues. Photogrammetric Engineering & Remote Sensing, 57(6),677-687. Lunetta, R.S., Ediriwickrema, J., Johnson, D.M., Lyon, J.G., & McKerrow, A. (2002). Impactsof vegetation dynamics on the identification of land-cover change in a biologicallycomplex community in North Carolina, USA. Remote Sens. Environ., 82, 258-270. Lunetta, R.L., Johnson, D.M., Lyon, J.G. & Crotwell, J. (2004). Impacts of imagery temporalfrequency on land-cover change detection monitoring. Remote Sens. Environ., 89(4),444-454. Lyon, J.G., Yuan, D., Lunetta, R.S., & Elvidge, C.D. (1998). A change detection experimentusing vegetation indices. Photogrammetric Engineering & Remote Sensing, 64(2), 143-150. MODIS (1999). MODIS Vegetation Index (MOD 13): Algorithm Theoretical Basis Document Page 26 of 29(version 3), (http://modis.gsfc.nasa.gov/data/atbd/atbd_mod13.pdf). (NASS) National Agriculture Statistical Service (2002). Census of Agriculture: Volume 1 Geographic Area Series. http://www.nass.usda.gov/census. Reed, B.C. (2006). Trend analysis of time-series phenology of north America derived from satellite data. GIScience & Remote Sensing, 43(1), 24-38. Richards, J.A. (1984). Thematic mapping from multitemporal image data using the principalcomponents transformation. Remote Sens. Environ., 16, 25-46. Roberts, H., Lehar, J., & Dreher, J.W. (1987). Time series analysis with CLEAN. I. Derivationof a spectrum. Astronomical Journal, 93(4), 968-989. Roerink, G.J., and Menenti, M. (2000). Reconstructing cloudfree NDVI composites usingFourier analysis of time series. Int. J. Remote Sensing, 21(9), 1911-1917. Sakamoto, T., Yokozawa, M., Toritani, H., Shibayama, M., Ishitsuka, N., and Oho, H. (2005). Acrop phenology detection method using time-series MODIS data. Remote Sens. Environ.,96, 366-374. Singh, A. (1989). Digital change detection techniques using remotely-sensed data. Int. J. Page 27 of 29Remote Sensing, 10(6), 989-1003. Townshend, J.R.G., and Justice, C.O. (1986). Analysis of the dynamics of African vegetation using the normalized difference vegetation index. Int. J. Remote Sensing, 8(8), 1189-