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. Level I Feature Extraction: Urban, Barren, Water, Forest, and Agriculture. 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) Transportation, (3) Water, (4) Forest, and (5) Agriculture. The second step involved separating Level 1 themes into more specific "Level 2" themes.

Landuse/Landcover Themes

The purpose of this levelized approach was to eliminate, as much as possible, potential error from the classification of agricultural land. All Level 2 (raster layer) categories were combined into 3 final landuse/landcover 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).

Urban Features

The Level I urban theme was extracted by identifying all areas within 120 meters of any road or railroad which fall within the boundaries of an incorporated urban area (U.S. Census Bureau TIGER incorporated areas). Those areas were treated as potential urban areas. Next, an ISOCLUSTERING classification algorithm under a potential urban area binary mask was performed on TM data from two image dates (bands 2,3,4,5, and 7 from Summer and Fall). The resulting clusters were aggregated, based on visual image interpretation of Fall TM scene bands 5,4,3 (RGB), into 3 categories: Level 1 Urban (Low Intensity), Level 2 Urban (Moderate Intensity), Level 3 Urban (High Intensity), Other Urban (Parks, Cemeteries, etc.), and Non-Urban. Non urban areas were subtracted from the potential urban mask to create a new binary urban mask. Finally, a logical NOT operation was performed on the urban mask to produce a binary mask that covered all non-urban areas.

Transportation Features

Level 1 and 2 Transportation theme comes directly from rasterized Arkansas Highway and Transportation Department (AHTD) and National Transportation Safety Board (NTSB) vector transportation data. Level 1 transportation is comprised of all Level 2 transportation themes: Major roads, railroads, and Airports. Major road and railroad data came from the AHTD, and airport data came from the NTSB. These vector layers were rasterized. All transportation pixels were subtracted from the non-urban area binary mask to create a new non-urban/non-transportation binary mask.

Water Features

Level I water was extracted by taking a ratio of bands 1 and 7 (B1/B7) for each scene date. Water pixels have relatively high values (DNs) in band 1 and the lowest DNs in band 7. Therefore the highest values resulting from the ratio corresponded exceedingly well with water. Water was extracted from the 1/7 ratio image using a simple grey-level threshold. For each scene footprint a permanent water category was created by comparing the water extent in each date and applying a logical AND operation. Anything that was water in all three scene dates was labeled permanent water. Correspondingly, water in any scene that was not a part of the permanent layer was considered flooded for the particular scene date: spring flooded, summer flooded, fall flooded.

Level 1 Forest, Barren, and Agriculture

Level 1 forest, barren, and agricultural features were extracted by applying an unsupervised clustering algorithm on TM bands 2, 3, 4, and 7 of the Autumn image. The resulting clusters were aggregated into one of three categories: forest, barren, agriculture. Barren contains only one Level 2 category, so Levels one and 2 are identical. The Level 1 agriculture scene varies from season to season due to the existence of flooded fields (from the water: flooded category). Flooded fields were subtracted from the Level 1 agriculture binary map for each season.

Level 2 Forest

Level 2 forest categories came directly from the Arkansas GAP analysis project's (2 hectare minimum mapping unit) vegetation map. All pixels beneath the forest binary bitmap were assigned the corresponding GAP category in the resulting Level 2 forest theme. Any pixels identified as non-forest by GAP that were in the forest binary map were labeled as forest-unclassified.

Level 2 Agriculture

(While all four county groups were processed in a similar manner, the specifics discussed below relate to group 2).

Spring

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 still 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

The differentiation and extraction of warm season crops was accomplished with a two step process. The first step was to perform an unsupervised isoclustering classification on TM bands 2,3,4,5, and 7 of the Summer and Fall scenes under the Level 1 agland mask. As noted above, 30% of the collected ground-truth data was to be used for classification purposes. A report was generated to compare the results of the unsupervised classification with the ground-truth binary maps. The report made it obvious that three Level 2 agriculture categories fell into well-defined cluster groups and were easily separated spectrally. All rice (grown in standing water), fallow vegetated, and fallow bare soil acreage was extracted with the initial unsupervised classification. The remaining agriculture categories were not as readily separable spectrally. Only three clusters, comprising 21% of the soybean binary map area, were clearly soybean. Similarly, only two clusters, comprising 33% of the cotton binary map, were clearly cotton. The figures were only 37% for sorghum and 26% for corn. These known clusters were aggregated into their respective categories. The aggregated image was examined for inconsistencies based on visual interpretation of the Summer TM scene bands 5,4,3 (RGB), a knowledge of cropping patterns, and crop phenology information. Upon examination, one cluster was removed from the rice aggregate. All aggregated areas were removed from the agriculture binary mask before step two: supervised classification. Several spectral scatterplots (variations of Summer and Fall TM bands 2, 3, 4, 5, and 7) were generated for the four remaining Level 2 categories: soybeans, cotton, sorghum, and corn using the ground-truth binary maps. It was found that the soybean scatterplots were bimodal in the Summer and Fall 3-4 and the Summer and Fall 2-4 combinations. When the cotton and soybean scatterplots were displayed in the same feature space, it was found that cotton coincided greatly with one end member of the bimodal soybean scatterplot. All crop phenologies, interview records, and field notes were carefully examined once again. With this information the Summer and Fall scenes were examined once again, it was discovered that the bimodal nature soybean scatterplot was directly related to the presence of early and late soybeans. Cotton was strongly correlated with early soybeans. Therefore, two non-parametric soybean signatures for Summer and Fall bands 2, 3, 4, 5, and 7 were created (drawn in feature space) using the 3-4 Summer scatterplot: early soybeans and late soybeans. Parametric signatures for cotton, sorghum, and corn for Summer and Fall bands 2, 3, 4, 5, and 7 for areas under the binary ground truth crop maps were also created . A maximum likelihood supervised classification was performed with mixed results. The separability of soybeans and cotton category had improved dramatically, but sorghum and cotton were still virtually inseparable. New signatures were created using only the Summer bands 2, 3, 4, 5, and 7. The results of the second maximum likelihood classification for sorghum and corn only were only slightly better, and it was decided that the corn and sorghum categories would be merged.

Fall

The extraction of Level 2 Fall agriculture categories was a straightforward process. Only two crops are in the ground in October: Cotton and Soybeans. In the southern portion of the study are the cotton harvest can extend well into mid November. Soybeans are harvested as late as mid December. All rice, sorghum, and corn has 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 agland mask. The resulting clusters were aggregated, based on visual image interpretation of Fall TM scene bands 5,4,3 (RGB), into 3 categories: fallow vegetated, 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).

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 11 pixels (approximately 2.5 acres) were merged into the category of its largest neighbor. The mode and sieve filters produced a map with more homogenous farm fields.