By Bruce Gorham, Research Scientist

Center for Advanced Spatial Technologies, University of Arkansas

ABSTRACT

Optimal use of soil and water resources is one of the principal challenges facing Arkansans.  Numerous water and soil conservation problems are directly related to land-use and land-cover, including surface and ground water pollution, topsoil loss, increasing soil salinity levels, and ground water depletion.  Different land-use and land-cover types have various impacts on soil and water resources. Therefore, it is important to track and map changes in land-use and land-cover.  Accurate land-use maps can help soil and water scientists to identify potential problem areas, predict where problems are likely to occur in the future, and to model appropriate solutions. The land-use and land-cover data generated for the project described herein is for the state of Arkansas in the year 1999.

In 1998 the Arkansas Soil and Water Conservation Commission (ASWCC) provided funding to the Center for Advanced Spatial Technologies at the University of Arkansas to develop 3-season digital land-use/land-cover maps focusing on agricultural land-use for the 75 counties of Arkansas. Moderate resolution satellite imagery: Landsat Thematic Mapper 5 (TM5) imagery was the chief data source for producing the digital LULC maps. Extensive ground-truth data was also collected to aid in the land-use/cover classification process. Combined with existing spatial data, the information produced from this project will serve as a basis for the formulation of water and soil resource management policies and practices.

Acknowledgments

The author wishes to thank the following people and for their hard work, kind cooperation, and useful advice. First, I would like to recognize the work of three high talented individuals: Linda Fye, Stephan Pollard, and Cindy Thatcher. Their hard work and dedication to the project ensured its success. I offer special thanks to Dr. Michael Garner. Mike assisted with the design of the crop ground-truth data collection methodology as well as the actual collection of that data in the field during the hot summer of 1999. To Jim Farley, thank you for your wise advice. Finally, to Noah Minard, and Alynne Bayard, thank you for helping out as the project was winding down to the deadline.

INTRODUCTION

Throughout the mid to late 1990's natural resource planners in Arkansas became increasingly concerned about a number of critical soil and water conservation issues within their state.  Once bountiful aquifers were in serious decline. The Mississippi Alluvial Aquifer and Sparta Aquifer, which together provide most of groundwater for cropland irrigation and aquaculture, were at dangerously low levels. They are still considered to be in "critical" condition by the Arkansas Soil and Water Conservation Commission (ASWCC).  Pollution of ground and surface water from point and non-point sources in both urban and rural areas continued to be a problem throughout our state.  Soil problems such as topsoil loss and surface salinization were of considerable concern to farmers and agriculturalists. The “Arkansas Water Resources and Wetlands Task Force” was created to address these problems.  The overall purpose of the task force was to "preserve and protect Arkansas water resources." One digital data set initially deemed essential to the aims of the task force was a map of agricultural land-use patterns.

In December 1996 the ASWCC funded the “Mississippi Alluvial Valley of Arkansas – Land-use/Land-cover Project” (MAVA-LULC) based on a proposal submitted by the Center for Advanced Spatial Technologies (CAST) at the University of Arkansas - Fayetteville. The proposal entitled "Development of a Digital land-use/land-cover Map for the Arkansas Delta" presented details on the development of enhanced digital land-use/land-cover maps focusing on agricultural crops for the twenty-seven Arkansas counties located within the Mississippi Alluvial Valley. The project “scope of work” stipulated that CAST would, with provided ASWCC funding, "Create a detailed, consistent land-use/land-cover map and a digital database for the 27 counties of the Delta." The 1992 project was successful, portraying agricultural land-use for that year with a high degree of accuracy.

In 1999 it was determined that an up-to-date land-use/land-cover map was needed. Once again, the Arkansas Soil and Water Conservation Commission provided funding for that mapping effort. The goal of the new LULC project was to map the land-use and land-cover for the entire state for the year 1999.  Like the 1992 dataset, the 1999 maps would be derived from Landsat Thematic Mapper 5 (TM5) satellite imagery. The new LULC maps built upon, and substantially improve the existing but incomplete digital LULC maps already available for the state as a whole. The project produced three digital maps (seasonal maps: spring, summer, and fall) focusing primarily on agricultural land-use.

STUDY AREA

The study area discussed herein is the entire state of Arkansas. While the satellite images employed in the study extended over state boundaries, only areas within the state were considered. The digital state boundary line used to clip the project data was produced by the Arkansas Highway and transportation Department.

Nearly all of Arkansas's harvested cropland acreage is devoted to six principal crops: soybeans, rice, cotton, wheat, sorghum, and corn.  The vast majority of Arkansas' harvested cropland is found in the eastern third of the state in 27 contiguous counties oriented north-to-south along the Mississippi River known as the Mississippi Alluvial Valley: “the Delta.” The soils of these regions are fertile and well suited to large-scale crop production.  Crop and pastureland is also found in the floodplains of the Arkansas River Valley in west-central Arkansas and the Red River Valley in southwest Arkansas. Crowley's Ridge, a three to 10 mile wide north-to-south erosional remnant, rises up to 50 meters above the fertile Mississippi Alluvial Valley in Northeast Arkansas. Crowleys Ridge is dotted with row crop fields, pastureland, and orchards.  The West Gulf Coastal Plain in southern Arkansas is characterized by an expanse of temperate mixed evergreen forests and mixed pastureland. The Ouachita uplands are predominately mixed-forest covered, but pastures are found in cleared valleys.  The Ozark highlands, like the Ouachitas, are forest-covered, but high quality pastureland is more common within the Ozark Plateaus.

PROCESS

GROUND-TRUTH DATA COLLECTION

The first phase of the project was the collection of ground-truth data (GT). The GT data was collected for two purposes: (1) to create multispectral signatures for supervised image classification, and (2) for accuracy assessment. A Trimble “Pro-XR” GPS unit was used to collect the GT samples for the following targets: Soybeans, Cotton, Rice, Sorghum/Corn, Warm Season Pasture, and Cool Season Pasture. (Note: GT was not collected for the winter wheat crop depicted in the Spring LULC map). The GT data collection was done over a four-week period in August and September. Week 1 was spent collecting GT crop information in southern Delta counties. Week 2 was spent collecting GT crop information in northern Delta counties. Week 3 was spent collecting cool vs. warm season pasture GT information in north and central Arkansas. Week 4 was spent collecting GT crop and pasture information in the Arkansas and Red River Valleys.  Upon completion of the GT data collection, the points were displayed on-screen over summertime Landsat satellite scenes. The fields corresponding to the GPS points were digitized as polygons and attributed with the appropriate category number and label. One-half of the resulting GT fields were used for signature generation. The remaining fields were reserved for accuracy assessment. Pasture categories were not extracted with a supervised classification scheme; therefore, the pasture GT points and polygons were used for accuracy assessment purposes only.

SATELLITE DATA ACQUISITION:

A total of ten Landsat 5 TM footprints are needed to cover the state of Arkansas: 23/35, 23/36, 23/37, 24/35, 24/36, 24/37, 25/35. 25/36, 25/37, and 26/35. Three scenes (spring, summer, and fall dates) were purchased for each footprint, making a total of 30 required scenes. In order to take advantage of maximum spectral variance between crops and pasture types for optimal scene selection, the planting patterns and phenologies of the major crops, as described by the Arkansas Agricultural Statistics Service (1998), were examined. Informal interviews with both FSA field office personnel and local University of Arkansas agronomy department faculty were also conducted for that purpose. With that information optimal crop and pasture identification target dates were identified for each season. 

Images were selected from the Landsat 5 TM online image archive maintained by the U.S. Geological Survey’s EROS Data Center archive based on three factors: overall image quality rating, percent cloud cover in image, and nearness to target date.  The thirty images based on the criteria described above were selected, and the order was placed with Earthsat, Inc. When received, the images were inspected for system error, as well as errors in cloud cover estimation. Some errors were found to exist in 3 scenes. They were returned and exchanged for other scene dates. The entire study area was covered for all of the following time periods: March/April, August, and September/October. 

DATA PREPROCESSING

The PCI Geomatics' GCPWorks module was used to register the 30 uncorrected TM scenes. Images were tied down to fit Arkansas Highway and Transportation Department road vector data (UTM projection, NAD83 datum).  A first order polynomial transformation model with nearest-neighbor resampling was used to rectify the image. The residual root mean square error was consistently less than one pixel on both the X and Y-axes. 

Upon inspecting the images for radiometric errors, several bright spots, flanked on both sides by rays of either maximum (255) or minimum (0) DN values were discovered in the mid IR bands of several scenes. The rays extended outwards horizontally as little as a few hundred meters to as much as several kilometers, and ranged from one to five pixels wide. It was discovered in the previous LULC study, that these fields were being burned as the sensor passed over.  These fires caused a calibration errors in TM bands 5 and 7.  To correct this radiometric error, a binary mask for the effected area was screen digitized, and an averaging filter with a 1 x 7 pixel kernel was run under the mask. Three iterations of the filtering process were needed to correct the calibration error. Some scenes were filtered to remove other minor system/calibration errors. No radiometric corrections were made to adjust for atmospheric conditions, or scene illumination (sun angle compensation). Finally, the nine scenes that overlap with other states were clipped so that all data outside Arkansas was removed from the file.

FEATURE EXTRACTION

For the purpose of this project we will define image classification as the extraction of differentiated classes or themes (land-use/land-cover information) from raw remotely sensed digital satellite data. The classification process for this project involved a two level approach. The first step in this approach was to separate the raw data into six broad LULC themes. These "Level 1" themes were (1) Urban, (2) Barren Land  (3) Water, (4) Herbaceous, and (5) Agriculture.  The second step involved separating Level 1 themes into more specific "Level 2" themes.  The intended purpose of this levelized approach was to eliminate, as much as possible, potential error from the classification of crop and pasture land.  All Level 2 (raster layer) categories were combined into 3 final land-use/land-cover maps: spring, summer, and fall. All urban, transportation, forest, barren, and perennial water categories remained the same for each season (permanent categories). Flooded and agricultural categories varied from season to season (variable categories).

LEVEL 1

The Level 1 stage was used to develop a generalized classification that identified forest, cropland/pasture, barren, herbaceous, water, and urban.  Water and urban were each classified separately, and the other categories were classified at the same time.  The same steps were used for each of the ten scenes and were as follows:

·         Perennial water was extracted by isolating low pixel values in the fall scenes.  An unsupervised classification was performed on the isolated pixels, which were classified as “water” or “not water”.  Results were then filtered and saved in a separate channel.

·         Urban areas were also extracted from fall scenes, but in a different manner.  A buffering operation was performed on urban roads, and a mask was created from the results.  This mask was combined with a second mask created from city-boundary polygons, resulting in one final mask that isolated potential urban areas.  An unsupervised classification was performed only on pixels located under this mask.  Pixels were placed in 1 of 4 categories: 1) not urban, 2) low intensity urban, 3) medium intensity urban, and 4) high intensity urban.  Results were filtered and placed in a separate channel.

·         The rest of the Level 1 categories were then classified using spring, summer, and fall seasons.  An unsupervised classification was executed on each entire scene, and pixels were aggregated into one of the following categories: 1) water 3 season,  2) water spring only, 3) water summer only, 4) water fall only, 5) water spring and summer, 6) water spring and fall, 7) water summer and fall, 8) barren (all seasons), 9) forest, 10) cropland/pasture, 11) herbaceous, and 12) urban features outside of urban areas.  This layer was left unfiltered.

·         Areas that were originally placed in one of the seven water categories were reexamined.  Those pixels that included non-perennial water (i.e., “spring only” or “summer and fall”) were isolated and reevaluated using an unsupervised classification with all 3 seasons to determine how those areas should be classified in seasons when they were not water.  For example, if a pixel was categorized as being water in spring and summer, then fall imagery was examined following the unsupervised classification. In these cases, pixels mainly fell into one of the categories of herbaceous, barren, crop/pasture, or forest.

·         The Level 1 categories that did not change from season to season, were written into 3 separate channels. The non-perennial water classifications were each combined with their corresponding Level 1 channels, so that the result was a separate final Level 1 classification for spring, summer, and fall.

·         The final Level 1 unsupervised classifications for all ten scenes for all three seasons were then mosaicked into three statewide files.

LEVEL 2

The Level 2 classification was the second major stage in the process, and involved performing a supervised classification on all areas identified as cropland/pasture during the Level 1 stage, and adding in forest classifications previously categorized in the 1992 Arkansas GAP project.  The steps were as follows:

• The first step involved preparing for the supervised classification process.  Six major physiographic regions exist in Arkansas with unique soil and topographic qualities and different major land use regimes.  These were used to develop masks that could help differentiate pasture from crop.  For example, agricultural areas in the Mississippi delta region would most likely be crops while areas in the Ozark Plateau would be pasture.  Some physiographic regions like the Arkansas River Valley, Crowley’s Ridge, and the Gulf Coastal Plain contain both pasture and crops. These had to be differentiated with an unsupervised classification.

• Using the physiographic region masks that isolated pasture, areas were then differentiated into cool- or warm-season pasture for each scene.  In general, cool-season pasture appears green in spring while warm-season pasture appears as bare soil.  The reverse is true in summer.  Normalized Difference Vegetation Index (NDVI) was calculated for both spring and summer.  The scene was then examined and thresholds for NDVI values (high when green, low when bare soil) were determined that represented pasture that was clearly warm- or cool-season.  Two masks were then developed by thresholding out these chosen values, one for warm-season pasture and one for cool-season pasture.  A maximum-likelihood supervised classification was then performed using the two masks as training sites, and the remaining pasture that was not easily distinguished by examination was classified accordingly.

• Spring Crops: The only crops grown during the spring season in the study area are wheat and oats.  Most of the land is being prepared for planting, or is weed covered.  It was found that Winter wheat and oats were confined to a fairly discrete range of DN values in TM band 4 of the March/April scene dates. A significant amount of other cool season grasses and legumes, forages often used for pasturage along the levees of the major rivers, were spectrally close to wheat and oats across all TM bands. These pastures, however, remained green throughout the growing season. By incorporating bands 3 and 4 of the  June TM coverage with the March/April coverage in an unsupervised classification it was possible to separate forage crops from forage pasture. Other aggregates from that unsupervised isoclustering algorithm included fallow vegetated (weedy) fields and bare soil/seedbeds.  All Level 1 and 2 spring map themes were then combined to create the Spring LULC map.

• Summer Crops: Under the physiographic region masks that isolated crop, areas were then distinguished using a maximum likelihood supervised classification.  Points that had been collected using a Global Positioning System (GPS) in the field during the summer and fall of 1999 were used as training sites.  Categories included: 1) soybeans, 2) rice, 3) cotton, 4) wheat/oats, 5) sorghum/corn, and 6) bare soil/seedbed.

• Fall Crops:  Fall agriculture categories were derived in a straightforward process. Only two crops are in the ground in October: Cotton and Soybeans.  In the southern portion of the study area, the cotton harvest can extend well into mid November. Soybeans are harvested as late as mid December.  All rice, sorghum, and corn had been harvested by mid October, and the empty fields are usually left bare. An unsupervised isoclustering algorithm (Fall TM bands 2, 3, 4, 5, and 7) was performed on all areas under the Level 1 agricultural land mask. The resulting clusters were aggregated, based on visual image interpretation of Fall TM scene bands 5,4,3 (RGB), into 3 categories: fallow/ bare, and cropped.  All pixels in the cropped aggregate were assigned either to a soybean or cotton category based on its summer category value (i.e. pixels that were cotton in the Summer and cropped in the Fall would be labeled cotton in the Fall map).

• All pasture and crop classifications were written into a state file over the original crop/pasture areas. 

• All forest, crop, barren, and herbaceous areas were then filtered for the entire state.

• Forest land cover classes are derived, in part, from the Arkansas Gap Project (Gap) (www.cast.uark.edu/gap/). The Forest Unclassified category reflects differences in image classification methodology between Gap and Arkansas LULC 1999. Most notably, large areas of the GAP urban residential category within municipal areas are here labeled as Forest Unclassified. This is because all pixels, including those within urban areas, were aggregated to the full range of categories for this study, while GAP placed urban pixels into one of its two urban categories only. Additionally, areas that had reforested between 1992 and 1999 are also labeled Forest Unclassified in this study. Deforested areas can be detected by comparing the 1992 GAP with the 1999 LULC data. For a complete description of the Gap classification scheme see Foti, T. et al., 1994.  “A class System for the Natural Vegetation of Arkansas”.  In Proceedings of the Arkansas Academy of Sciences, 48: 50-62.

CLASSIFICATION SCHEME

Level I                         Level II

Urban                          Intensity 1 (Low Intensity)

                                    Intensity 2 (Moderate Intensity)

                                    Intensity 3 (High Intensity)

                                    Other Urban (Parks, Cemeteries, Military Bases, etc.)

Water

                                    Perennial Water

                                    Flooded

Barren Land

                                    Sand, Rock Outcrops, Mining Operations

Herbaceous

                                    Herbaceous

Forest (GAP)

                                    Forest Unclassified

                                    Shortleaf Pine

                                    Loblolly Pine

                                    Eastern Red Cedar

                                    White Oak, Northern Red Oak, Shortleaf Pine, Hickory

                                    Loblolly Pine, Shortleaf Pine, Oak

                                    Eastern Red Cedar, Mixed Pine-Hardwood

                                    American Beech

                                    White Oak, Mixed Hardwoods

                                    Northern Red Oak, Mixed Oak

                                    Southern Red Oak, Mixed Oak

                                    Post Oak

                                    Eastern Red Cedar, Oak

                                    Shortleaf Pine, Oak

                                    White Cedar, Oak

                                    Oak, Black Hickory

                                    Overcup Oak (Quercus Lyrata)

                                    Water Hickory (Carya Aquatica)

                                    Cherrybark Oak (Quercus Falcata var. Pagodifolia)

Sugarberry (Celtis Laevigata)

Nuttall Oak

Willow Oak

Sweetgum (Liquidambar Styraciflua)

Baldcypress, Mixed Hardwoods

Baldcypress

Tupelo Gum (Nyssa)

Willow, Cottonwood

Birch, Sycamore, Maple

Agriculture

                                    Soybeans

                                    Rice

                                    Cotton

                                    Wheat/Oats

                                    Sorghum/Corn

                                    Bare Soil/Seedbed/Fallow

                                    Warm Season Pasture

                                    Cool Season Pasture

POST CLASSIFICATION PROCESSING

After all Level 2 themes were extracted, each seasonal map was constructed by combining all themes for that season into one image.  Each seasonal image was examined for theme continuity by farm-field.  Small pixels clumps (stray pixels) of one crop category were found within large pixel groupings (fields).  In order to make fields more homogenous a 3x3 mode filter was performed on only the crop categories of all three season maps. Next, using a "sieve" filter, all Level 2 agriculture pixel clumps smaller than 20 pixels (approximately 5 acres) were merged into the category of its largest neighbor. The mode and sieve filters produced a map with homogenous farm fields.

RESULTS

ACCURACY

Overall, the crop accuracies achieved in this study exceeded those of all previous crop mapping studies conducted in Arkansas, including those for the 1992 MAVA-LULC study. As noted above 50% of the field data GT data was reserved for accuracy assessment. Since ground truth was only collected for crops and pasture there was no way to quantify the accuracies of non-agricultural categories. Accuracy for spring map product (winter wheat/oats) was not performed because GT data was not collected for the spring crops. The summer crop accuracy report is in the form of a confusion matrix (see Table 1).

Additional Accuracy Information:

Average accuracy = 87.69%

Overall accuracy = 90.85%

Kappa Coefficient  =  0.88761   

Standard Deviation =  0.00068

Confidence Level :

99%  0.88761 +/- 0.00175

95%  0.88761 +/- 0.00133

90%  0.88761 +/- 0.00112

DIGITAL MAP RESULTS:  PIE CHARTS

DATA USES

The data products generated for the project are not intended for site specific planning or research. They were designed primarily for planning and research at the county, regional, or statewide level.  The 1999 LULC datasets should be of particular interest to those studying land-use change or land-cover changes. This dataset fits well with the 1992 LULC data for the 27 counties of the Mississippi of Arkansas with only minimal reclassification.

CONCLUSIONS

Overall, the results of the study are quite positive. The dataset produced for this project has several deficiencies mostly dealing with issues of resolution.  Increased frequency in temporal resolution offers the most promise for improving agricultural studies.  An increase in spectral resolution: hyperspectral sensors, or the addition of thermal IR data, will also improve future agricultural land-use mapping projects. Fine spectral detail will allow analysts to detect subtle differences between and within various crop types, making it feasible to conduct large-scale studies of crop varieties.

Spatial resolution was not a major factor in this study. While the new high resolution sensors may benefit precision agriculture studies, the high cost of the data and tremendous disk space requirements are currently cost prohibitive for large scale mapping projects. TM 7's improved thermal band resolution, however, will benefit future studies of the type conducted here, and TM 7's new 15 meter resolution panchromatic band may help to better delineate field boundaries at a considerably lower cost than SPOT imagery.

In the past most agricultural land-use maps portrayed a somewhat static, year-to-year, picture of the landscape. The maps generated for the MAVA-LULC project depict season-to-season land-use/land-cover patterns. Combined with other natural resource and socio-economic data, the information discussed here should prove to be a very useful information base for natural resource planners.