Elsevier

Remote Sensing of Environment

Volume 112, Issue 3, 18 March 2008, Pages 1096-1116
Remote Sensing of Environment

Large-area crop mapping using time-series MODIS 250 m NDVI data: An assessment for the U.S. Central Great Plains

https://doi.org/10.1016/j.rse.2007.07.019Get rights and content

Abstract

Improved and up-to-date land use/land cover (LULC) data sets that classify specific crop types and associated land use practices are needed over intensively cropped regions such as the U.S. Central Great Plains, to support science and policy applications focused on understanding the role and response of the agricultural sector to environmental change issues. The Moderate Resolution Imaging Spectroradiometer (MODIS) holds considerable promise for detailed, large-area crop-related LULC mapping in this region given its global coverage, unique combination of spatial, spectral, and temporal resolutions, and the cost-free status of its data. The objective of this research was to evaluate the applicability of time-series MODIS 250 m normalized difference vegetation index (NDVI) data for large-area crop-related LULC mapping over the U.S. Central Great Plains. A hierarchical crop mapping protocol, which applied a decision tree classifier to multi-temporal NDVI data collected over the growing season, was tested for the state of Kansas. The hierarchical classification approach produced a series of four crop-related LULC maps that progressively classified: 1) crop/non-crop, 2) general crop types (alfalfa, summer crops, winter wheat, and fallow), 3) specific summer crop types (corn, sorghum, and soybeans), and 4) irrigated/non-irrigated crops. A series of quantitative and qualitative assessments were made at the state and sub-state levels to evaluate the overall map quality and highlight areas of misclassification for each map.

The series of MODIS NDVI-derived crop maps generally had classification accuracies greater than 80%. Overall accuracies ranged from 94% for the general crop map to 84% for the summer crop map. The state-level crop patterns classified in the maps were consistent with the general cropping patterns across Kansas. The classified crop areas were usually within 1–5% of the USDA reported crop area for most classes. Sub-state comparisons found the areal discrepancies for most classes to be relatively minor throughout the state. In eastern Kansas, some small cropland areas could not be resolved at MODIS' 250 m resolution and led to an underclassification of cropland in the crop/non-crop map, which was propagated to the subsequent crop classifications. Notable regional areal differences in crop area were also found for a few selected crop classes and locations that were related to climate factors (i.e., omission of marginal, dryland cropped areas and the underclassification of irrigated crops in western Kansas), localized precipitation patterns (overclassification of irrigated crops in northeast Kansas), and specific cropping practices (double cropping in southeast Kansas).

Introduction

Georeferenced land use/land cover (LULC) data sets are primary inputs for environmental modeling and monitoring, natural resource management, and policy development. A variety of LULC data sets are needed to support the growing and diverse demands of the global environmental change research community (Cihlar, 2000, DeFries and Belward, 2000). Several major research programs and documents (NASA, 2002, NRC, 2001, Sarmiento and Wofsy, 1999, Turner et al., 1995) have identified the development of improved and up-to-date regional- to global-scale LULC products as a research priority. These products should characterize current LULC patterns, document major LULC changes, and place more emphasis on land use in the thematic classification schemes of the maps.

Over the past decade, the science of large-area LULC mapping has made considerable strides as remotely sensed data and computing resources have improved and advanced classification techniques have emerged (DeFries & Belward, 2000). During this period, large-area LULC mapping has evolved through numerous efforts at state (Craig, 2001, Eve and Merchant, 1998), regional (Bosard et al., 2000, Homer et al., 2004, Vogelmann et al., 2001), and global (Bartholome and Belward, 2005, DeFries et al., 1998, DeFries and Townshend, 1994, Friedl et al., 2002, Hansen et al., 2000, Loveland et al., 2000) scales. In addition, considerable effort has been made to advance large-area LULC characterization beyond the traditional thematic classification of specific LULC types by mapping continuous land cover fields (Hansen et al., 2002), land cover change (Justice et al., 2002, Zhan et al., 2002), and biophysical land cover characteristics such as leaf area index (LAI) and fraction of absorbed photosynthetically active radiation (FPAR) (Myneni et al., 2002).

Despite this progress, little emphasis has been placed on large-area crop mapping and monitoring. Most of the mapping efforts highlighted above have focused on the classification of land cover types associated with natural systems (e.g., forest, grassland, and shrubland) and have tended to generalize cropland areas into a single or limited number of thematic classes. Few large-area mapping efforts have attempted to map specific crop types and associated land use practices (Craig, 2001), particularly on the short time-step that is required to reflect the rapid land cover changes that commonly occur from year to year in crop rotations.

Timely crop-related LULC information is currently limited over major agricultural regions such as the U.S. Central Great Plains, which face a number of environmental issues (e.g., climate change and groundwater depletion) that threaten the area's long-term sustainability (Ojima & Lackett, 2002). Cropland areas are intensively managed and modified through a variety of human activities, which can have a wide range of impacts on biogeochemical and hydrologic cycles, climate, ecosystem functions, the economy, and human health. As a result, new mapping protocols are needed to characterize the regional-scale cropping patterns on a repetitive basis and provide improved LULC information to scientists and policy makers.

The objective of this study was to investigate the applicability of time-series Moderate Resolution Imaging Spectroradiometer (MODIS) 250 m normalized difference vegetation index (NDVI) data for regional-scale crop mapping in the U.S. Central Great Plains. A crop mapping methodology, which applied a decision tree classification technique to a time series of MODIS 250 m VI data spanning one growing season, was tested over the state of Kansas. A four-level hierarchical classification scheme was implemented, which produced a series of four crop-related LULC maps that progressively classified cropland areas into more thematically-detailed classes (Fig. 1). Three primary research questions were addressed in this study. First, what thematic accuracy can be achieved for classifying crop/non-cropland, general crop types (alfalfa, summer crops, winter wheat, and fallow), specific summer crop types (corn, sorghum, and soybeans), and irrigated/non-irrigated crops using a time series of MODIS 250 m NDVI data collected across the growing season? Second, are the general cropping patterns depicted in the series of 250 m maps consistent with the patterns reported across Kansas? Third, do the maps exhibit any regional trends or major areal deviations from the general cropping patterns reported for Kansas? Major regional differences may signal a limitation of the classification methodology in certain parts of the state or for specific crop-related LULC classes.

The development of a regional-scale crop mapping methodology is challenging because it requires remotely sensed data that have large geographic coverage, high temporal resolution, adequate spatial resolution relative to the typical field size, and minimal cost. Remotely sensed data from traditional sources such as the Landsat Thematic Mapper (TM and ETM+) and the Advanced Very High Resolution Radiometer (AVHRR) have proved useful for LULC characterization, but are limited for such an approach because of resolution limitations, data availability/quality issues, and/or data costs.

Most crop mapping using high resolution, multi-spectral data from Landsat TM/ETM+ (and similar sensors such as SPOT) has been conducted at a local scale (Mosiman, 2003, Price et al., 1997). The exception is the cropland data layer (CDL), which is a state-level classification of specific crop types produced by the United States Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) that is annually updated (Craig, 2001). However, the CDL is only available for a limited number of states (10 states in 2005) (NASS, 2006). Other large-area Landsat-based LULC mapping efforts such as the National Land Cover Dataset (NLCD) (Homer et al., 2004, Vogelmann et al., 2001) and Gap Analysis Program (GAP) state-level LULC maps (Eve & Merchant, 1998) have typically classified cropland into a limited number of generalized classes and have updated the maps infrequently. Large-area mapping using Landsat data has been limited by the considerable costs of acquiring and processing the large data volumes that are required (DeFries & Belward, 2000). A multi-seasonal Landsat classification approach is typically adopted and has consistently produced higher accuracies than a single-date approach for both general LULC (Wardlow & Egbert, 2003) and crop type (Price et al., 1997) mapping. As a result, a large number of Landsat scene-date combinations have to be processed to provide adequate spatial and temporal coverage, which is costly and time and labor intensive. The acquisition of relatively cloud-free imagery at optimal times during the growing season is also an issue because of the long revisit time of the Landsat instrument (DeFries & Belward, 2000).

Coarse resolution (1 km and 8 km), time-series AVHRR NDVI data have been widely used to classify continental- to global-scale land cover patterns (DeFries et al., 1998, DeFries and Townshend, 1994, Hansen et al., 2000, Loveland et al., 1995, Loveland and Belward, 1997). The high temporal resolution (10- to 14-day composites generated from near-daily image acquisitions) of the AVHRR time-series data in combination with the NDVI's strong relationship with biophysical vegetation characteristics such as LAI and green biomass (Asrar et al., 1989, Baret and Guyot, 1991) enables land cover types to be discriminated based on their unique phenological responses. However, the coarse spatial resolution of AVHRR data limits both the thematic and spatial detail of the LULC types that can be mapped. Most AVHRR pixels have an integrated spectral–temporal signal from multiple LULC types contained within the 1-km footprint (Townshend and Justice, 1988, Zhan et al., 2002). As a result, coarse resolution imagery is suitable for classifying the general, broad-scale patterns of natural systems, but the high spatial variability and complexity of agricultural systems requires higher resolution data (Turner et al., 1995). The use of 1-km AVHRR data for LULC classification in agricultural regions is difficult (Loveland et al., 2000) and has produced inconsistent results (Loveland et al., 1995). Most AVHRR-related mapping efforts have characterized broad-scale land cover patterns, classified general vegetation types (e.g., deciduous, broadleaf forest) and/or aggregated land cover classes (e.g., crop/grassland mosaic), and assigned cropland areas to either a single or mixed crop/natural vegetation class.

The MODIS instrument offers new possibilities for large-area crop mapping by providing a near-daily global coverage of science-quality, intermediate resolution (250 m) data since February 2000 at no cost to the end user (Justice & Townshend, 2002). MODIS 250 m VI data, in particular, are well suited for this type of application in the U.S. Central Great Plains. The calibration, atmospheric correction (water vapor and aerosols) (Vermote et al., 2002), and relatively high sub-pixel geolocational accuracy (∼ 50 m (1σ) at nadir) (Wolfe et al., 2002) of MODIS data enable distinct multi-temporal VI signals of specific crops to be detected at the 250 m pixel level (Wardlow et al., 2006, Wardlow et al., 2007). In an analysis of 16-day composited MODIS 250 m VI data for 2000+ fields in Kansas, Wardlow et al. (2007) found the data to have sufficient spatial, spectral, temporal, and radiometric resolutions to detect unique multi-temporal VI signatures for the state's major crop types (Fig. 2) and land use practices (Fig. 3). Alfalfa, summer crops (corn, sorghum, and soybeans), winter wheat, and fallow (unplanted, idle fields) were highly separable in the VI data at some point during the growing season because of their very distinct crop calendars. Specific summer crops such as corn, sorghum, and soybeans were less separable, but subtle differences in their specific planting times and general growth patterns were captured in the VI data, which can be used to discriminate these crop types. Irrigated and non-irrigated fields of the same crop type also exhibited different VI responses, with irrigated crops maintaining higher VI values across most of the growing season as demonstrated by corn and winter wheat in Fig. 3.

The moderate 250 m spatial resolution of MODIS is appropriate for classifying cropping patterns in the U.S. Central Great Plains given the region's relatively large field sizes, which are frequently 32.4 ha or larger and would spatially correspond to five or more 250 m pixels. The two 250 m spectral bands (620–670 nm and 841–876 nm) on MODIS used to calculate the VI data were intended to be used to detect anthropogenic-driven land cover changes (Townshend & Justice, 1988), and the value of data at this spatial scale has already been demonstrated for LULC change detection (Lunetta et al., 2006, Morton et al., 2006, Zhan et al., 2002), continuous fields land cover mapping (Hansen et al., 2002, Lobell and Asner, 2004), general land cover mapping (Wessels et al., 2004), and vegetation phenology characterization (Wardlow et al., 2006). Wardlow et al. (2007) noted that similar landscape-level cropping patterns could be visually discerned in MODIS 250 m and Landsat ETM+ 30 m imagery throughout most of Kansas. Hansen et al. (2002) also reported that other landscape features associated with human activity (e.g., deforestation) were visible at MODIS' 250 m spatial resolution. The utility of the MODIS 250 m VI data sets for several cropland characterization activities has been demonstrated (Lobell and Asner, 2004, Morton et al., 2006, Wardlow et al., 2006), but their suitability for mapping specific crop types and crop-related land use practices has yet to be fully explored.

Large-area LULC mapping has improved with the application of advanced classification techniques such as decision tree (DT) classifiers, which have several advantages over traditional supervised classifiers (Hansen et al., 1996) and have consistently produced higher classification accuracies for this task (Friedl and Brodley, 1997, Hansen et al., 1996). DTs are non-parametric and can handle multi-modal distributions in the input data because they operate on thresholds in multi-spectral space rather than measures of central tendency (Hansen et al., 2000). This is critical for regional-scale crop mapping because of the considerable intra-class variability exhibited in the time-series MODIS NDVI data for a given crop due to regional variations in climate and management practices (Wardlow et al., 2007, Wardlow et al., 2006). DTs can also handle non-linear and hierarchical relationships between the input variables and the classes, as well as a variety of data types. They are also efficient at processing the large data volumes that are required for large-area applications. As a result, DT techniques are increasingly being used for large-area LULC mapping with success (Friedl et al., 2002, Homer et al., 2004, Wessels et al., 2004). The use of other techniques from the machine learning community such as boosting (Freund, 1995, Shapire, 1990) and bagging (Bauer and Kohavi, 1999, Breiman, 1996) in combination with DTs has also further improved LULC classification capabilities (Friedl et al., 1999, Lawrence et al., 2004).

Section snippets

Study area

Kansas is an agriculturally-dominated state located in the U.S. Central Great Plains (Fig. 4) with 46.9% (10.0 million ha) of its total area dedicated to crop production. The state's major crops include alfalfa (Medicago sativa), corn (Zea mays), sorghum (Sorghum bicolor), soybeans (Glycine max), and winter wheat (Triticum aestivum). Over the past decade, Kansas has ranked among the top 10 states in both acreage and production for most of these crops and has, on average, $3.3 billion in total

MODIS 250 m NDVI data

A 15-date time series of MODIS 250 m NDVI data (MOD13Q1, Collection 4) spanning from the March 22 to November 1 composite periods was created for Kansas. Data were required from three MODIS tiles (h09v05, h10v05, and h10v04) for statewide coverage. The tiled NDVI data were mosaicked, reprojected from the Sinusoidal to Lambert Azimuthal Equal Area (LAEA) projection, and subset over Kansas for each composite period and then sequentially stacked to produce the time-series data set.

Field site training and validation database

A database of

Hierarchical classification scheme

A four-level, hierarchical classification scheme (Fig. 1) was implemented in this study. At the initial stage (Level 1), the entire study area was mapped into crop and non-crop classes. The cropland areas were then mapped into four general crop types (Level 2). The summer crop areas classified in the general crop map were then further classified into three specific summer crop types (Level 3). The final step was to classify the cropland area into irrigated and non-irrigated classes (Level 4).

Visual assessment

The major crop/non-crop patterns in the MODIS-derived map (Fig. 6a) were consistent with Kansas' general land cover patterns. Major crop areas such as the Western Corn Belt (northern ASD 70), the Winter Wheat Belt (southeast ASD 60), and the large cropland expanses throughout the central and western ASDs were depicted. The cropped floodplain areas and smaller, fragmented crop patches throughout the eastern ASDs (80 and 90) were classified. Major non-crop features such as the Flint Hills

Conclusions

This study has demonstrated that time-series MODIS 250 m NDVI data provide a viable option for regional-scale crop mapping in the U.S. Central Great Plains. Relatively high classification accuracies (generally > 84%) were attained across the series of crop-related LULC maps produced for Kansas, and the classified crop patterns were consistent with the reported cropping patterns and crop distributions for the state. The classification accuracies were comparable to those of previous Landsat-based

Acknowledgements

This research was supported by NASA Headquarters under the Earth System Science Fellowship Grant NGT5-50421 and the USGS AmericaView program. The work was conducted at the Kansas Applied Remote Sensing (KARS) Program (Edward A. Martinko, Director). The authors thank the Kansas Farm Service Agency and 48 of its county offices for providing the annotated aerial photographs used for field site selection. The authors also thank Dr. Thomas Loveland and Dr. Bruce Wylie of the USGS Center for Earth

References (73)

  • R.S. Lunetta et al.

    Land-cover change detection using multi-temporal MODIS NDVI data

    Remote Sensing of Environment

    (2006)
  • R.B. Myneni et al.

    Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data

    Remote Sensing of Environment

    (2002)
  • S.V. Stehman et al.

    Design and analysis for thematic map accuracy assessment: Fundamental principles

    Remote Sensing of Environment

    (1998)
  • E.F. Vermote et al.

    Atmospheric correction of MODIS data in the visible and middle infrared: First results

    Remote Sensing of Environment

    (2002)
  • B.D. Wardlow et al.

    Analysis of time-series MODIS 250 m vegetation index data for crop classification in the U.S. Central Great Plains

    Remote Sensing of Environment

    (2007)
  • K.J. Wessels et al.

    Mapping regional land cover with MODIS data for biological conservation: Examples from the Great Yellowstone Ecosystem, USA and Para State, Brazil

    Remote Sensing of Environment

    (2004)
  • R.E. Wolfe et al.

    Achieving sub-pixel geolocation accuracy in support of MODIS land science

    Remote Sensing of Environment

    (2002)
  • X. Zhan et al.

    Detection of land cover changes using MODIS 250 m data

    Remote Sensing of Environment

    (2002)
  • G. Asrar et al.

    Estimation of plant canopy attributes from spectral reflectance measurements, Chapter 7

  • E. Bartholome et al.

    GLC2000: A new approach to global land cover mapping from Earth observation data

    International Journal of Remote Sensing

    (2005)
  • E. Bauer et al.

    An empirical comparison of voting classification algorithms: Bagging, boosting, and variants

    Machine Learning

    (1999)
  • Bosard, M., Feranec, J., & Otahel, J. (2000). CORINE Land Cover Technical Guide — Addendum 2000. European Environmental...
  • L. Breiman

    Bagging procedures

    Machine Learning

    (1996)
  • J.C. Brown et al.

    Multi-temporal, moderate spatial resolution remote sensing of modern agricultural production and land modification in the Brazilian Amazon

    GIScience and Remote Sensing

    (2007)
  • I.C. Burke et al.

    Regional analysis of the Central Great Plains

    Bioscience

    (1991)
  • J. Cihlar

    Land cover mapping of large areas from satellites: Status and research priorities

    International Journal of Remote Sensing

    (2000)
  • M.E. Craig

    The NASS cropland data layer program

  • R.S. DeFries et al.

    Global and regional land cover characterization from satellite data: An introduction to the special issue

    International Journal of Remote Sensing

    (2000)
  • R.S. DeFries et al.

    NDVI-derived land cover classifications at a global scale

    International Journal of Remote Sensing

    (1994)
  • R.S. DeFries et al.

    Global land cover classifications at 8 km spatial resolution: The use of training data derived from Landsat imagery in decision tree classifiers

    International Journal of Remote Sensing

    (1998)
  • R.E. Dickinson et al.

    Biosphere–atmosphere transfer scheme (BATS) version as coupled to the NCAR community climate model

  • S.L. Egbert et al.

    The Kansas Gap Land Cover Map: Final Report, Kansas Biological Survey Report #98

    (2001)
  • M.D. Eve et al.

    National Survey of Land Cover Mapping Protocols Used in the GAP Analysis Program

  • R.T.T. Forman et al.
  • M.A. Friedl et al.

    Maximizing the land cover classification accuracies produced using decision trees at continental to global scales

    IEEE Transactions on Geoscience and Remote Sensing

    (1999)
  • M.C. Hansen et al.

    Global land cover classification at 1 km spatial resolution using a classification tree approach

    International Journal of Remote Sensing

    (2000)
  • Cited by (0)

    View full text