The information on data quality was generated by the Decision Tree algorithm that conducts a cross-validation for assessing classification and prediction reliability. No formal independent accuracy assessment of the land cover has been made. The regression tree algorithm employed in mapping the land cover offers a cross-validation option for assessing classification and prediction reliability. Cross-validation can provide relatively reliable estimates for land cover predictions if the reference data used for cross-validation are collected based on a statistical valid sampling design. For the land cover modeling, a 10-fold cross-validation was conducted by dividing the entire training data set into 10 subsets of equal size. For each model run, an accuracy estimate was derived using one subset to evaluate the model prediction (with the model developed using the remaining training samples). This process was repeated 10 times. After all 10 runs, an average value of all accuracy estimates from the 10 runs were computed. Users should be cautioned that these cross-validation results provide users with only first-order estimates of data quality, and should not be considered a formal accuracy assessment.
The 2005 GLUT land cover is a satellite imagery derived map of generally coarse nature; it is not intended to be used at scales finer than approximately 1:100,000. The accuracy of the map varies by class, and users interested in particular classes should consult tables listing specific class accuracies. There are rarely sharp lines delineating natural land cover types in Georgia. In Georgia, we mapped areas of open ocean seven miles out from the coast line. All land cover figures include these areas of open ocean.
Additional information may be found at <http://narsal.ecology.uga.edu>
The primary data sets used were mosaicked +ETM (TM imagery) satellite imagery from 2005 for spring, leafoff, and leafon conditions. Also used were TM derived tasseled caps, TM thermal bands, digital elevation models (DEMs), DEM slope calculations, and selected NAPP DOQQs (used for training sets).
Selection of training data for Georgia consisted of identifying areas that have remained unchaged since 2001. This was accomplished by creating a conservative change detection with GLUT 2001 TM imagery and 2005 TM imagery. Training data was then sampled from those areas that were determined to be unchanged and identified as the corresponding GLUT 2001 Land Cover class.
Our target sample size, based on EDC recommendations, was approximately 3,000 training sample points. This was stratified by the sizes of the each class within the unchanged areas.
The inclusion of several classes relied on the use of ancillary data and hand digitizing to incorporate them into the database. The demarcation line between freshwater and brackish wetlands was determined using the NWI map.
Note that the training data were used to map all land cover classes except for the two classes in urban and sub-urban areas. All urban and suburban land cover classes were mapped and quality assessed separately through a sub-pixel quantification of impervious surfaces using a regression tree modeling method.
Following the development of the best classification through decision tree modeling, additional steps were required to complete the final land cover product. The four classes in urban and suburban areas were determined from the percent imperviousness mapping product. The multiple hierarchical classifications were combined with the urban classes to complete the land cover product. Finally visual inspection of the classification was made with areas/pixels that were wrongly classified, such as cloud and shadow, were reclassified by using CART models which excluded the offending imagery.