A note about data scale:
Scale is an important factor in data usage. Certain scale datasets are not suitable for some project, analysis, or modelling purposes. Please be sure you are using the best available data.
1:24000 scale datasets are recommended for projects that are at the county level. 1:24000 data should NOT be used for high accuracy base mapping such as property parcel boundaries.
1:100000 scale datasets are recommended for projects that are at the multi-county or regional level. 1:125000 scale datasets are recommended for projects that are at the regional or state level or larger.
Vector datasets with no defined scale or accuracy should be considered suspect. Make sure you are familiar with your data before using it for projects or analysis. Every effort has been made to supply the user with data documentation. For additional information, see the References section and the Data Source Contact section of this documentation. For more information regarding scale and accuracy, see our webpage at: <http://geoplan.ufl.edu/education.html>
1. Unsupervised classifications were performed on each entire Landsat scene. Initial classifications were performed on all six 30 m pixel spectral bands. The number of resultant spectral classes was typically set to 75-100.
2. The 75-100 spectral classes resulting from Step 1 were reviewed individually. Each spectral class was visually checked against the Landsat imagery as well as the ancillary data. If any of the spectral classes consistently identified a specific target land cover type (e.g., mangrove swamp, pine forest, coastal strand), those spectral classes were labeled according to the vegetation or land cover type they represented, and those classes were considered final and were excluded from further analyses.
3. All unlabeled pixels remaining after Step 2 were then subjected to additional unsupervised classifications. Differing band combinations (i.e., subsets) often were used to group similar areas to a distinct cover type. Resultant spectral classes varied from a few to over 50. At this point the process became iterative, and these steps were repeated until all pixels fell into a specific land cover type or into a larger, temporary grouping (e.g., disturbed). Additionally, areas with unique features or areas resulting in classification "confusion" would be clipped from the scene. Unsupervised classification would then be performed only on the clipped areas.
4. The data sets resulting from Step 3 that consistently represented a specific natural land cover type were assigned the appropriate label, were added to the final data set, and were excluded from further analyses.
5. Agricultural and urban land use classes from the 1995 digital data set of statewide land use/land cover were then used as an overlay. Spectral classes that had been identified as disturbed and that fell within the agricultural or urban land use class overlay were isolated. Unsupervised classification was performed on these areas to spectrally isolate agricultural areas from urban areas.
6. By comparing the spectral classes resulting from Step 5 with the ancillary data sets (i.e., 1995 land use/land cover, 1999 DOQQs), disturbed spectral classes were categorized into six agricultural land use classes (i.e., improved pasture, unimproved pasture, sugar cane, citrus, row and field crops, other agriculture), two urban classes (i.e., high density urban, low density urban), and extractive (i.e., mining). All pixels in these classes were added to the final data set and were excluded from further analyses. Visual interpretation of the spectral classes and the Landsat imagery was often required in areas where there was new urban growth and where agricultural lands were in a bare soil state, creating a false urban signature. Very often it was necessary to isolate these areas individually and assign the appropriate label. Areas that classified as disturbed but were not within the agricultural and urban lands overlay were checked visually against the Landsat imagery and other ancillary data layers. Often these disturbed areas were new areas of agriculture or urban lands, or they represented recent land clearings due to silvicultural practices or other unknown causes.
7. Once an entire scene had been analyzed in the above manner, the biologist then examined specific geographic areas of similar physiographic features (e.g., coastal wetlands, xeric ridges), and, if necessary, performed additional unsupervised classifications on any remaining classes of pixels that could not be separated based on spectral information developed at the level of the entire Landsat scene. Any classes that consistently represented a specific land cover type were assigned the appropriate land cover label, added to the final data set, and excluded from further analyses.
8. Any remaining areas that did not have a specific land cover label were visually reviewed in relationship to the Landsat imagery, land use/land cover data, and DOQQs. If possible, unlabeled groups of pixels were assigned to appropriate land cover types by hand, and were added to the final data set and excluded from further analyses.
9. Once all pixels within a Landsat scene had been classified, labeled, and added to the final data set comprising the updated vegetation and land cover map, specific areas of the map were visited in the field for ground-truthing. Any mistakes discovered in the ground-truthing process were then corrected to create a final draft vegetation map covering the entire Landsat scene.
10. The final draft vegetation and land cover map for each scene was then reviewed by the project manager. The project manager compared each draft map against ancillary data sets and identified specific problem areas that either needed checking for accuracy or correction. Project manager recommendations were then returned to staff to make corrections needed to produce a final vegetation and land cover data set for each Landsat scene.
11. Early in the project, a number of the Landsat scenes purchased from EROS Data Center were from 2000-2002, and final drafts of vegetation and land cover for these scenes were based on these earlier dates. However, as luck would have it, 2003 was a good year for cloud-free satellite imagery in Florida. Thus, not only were the later scenes in the project mapped using only 2003 imagery, but also new 2003 Landsat ETM+ imagery was purchased for the entire state, and the new imagery was used to update disturbed areas of all earlier scenes to 2003 according to the following procedure.
a. Unsupervised classifications were conducted for an entire 2003 scene. b. Spectral classes representing sparsely vegetated areas (e.g., disturbed areas) were isolated. c. Disturbed areas from the 2003 imagery that were classified as natural vegetation in the earlier imagery (2000-2002) were isolated and further examined. d.The areas of new disturbance were then classified into appropriate categories. e. Additionally, other changes between the two scenes were examined and updated if necessary. f. All changes and updates between the two scene dates were then incorporated into the previously classified map to produce a new vegetation and land cover data set for each scene that reflected conditions in 2003.
12. Once a scene was complete and updated, if necessary it was edge-matched and merged with adjacent scenes that had previously been completed. Upon completion of last scene, all scenes were then merged, forming a single statewide map.