Impacts of vegetation dynamics on the identification of land-cover change in a biologically complex community in North Carolina, USA

A land-cover (LC) change detection experiment was performed in the biologically complex landscape of the Neuse River Basin (NRB), North Carolina using Landsat 5 and 7 imagery collected in May of 1993 and 2000. Methods included pixel-wise Normalized Difference Vegetation Index (NDVI) and Multiband Image Difference (MID) techniques. The NDVI method utilized non-normalized (raw) imagery data, while the MID method required normalized imagery. Image normalization techniques included both automatic scattergram-controlled regression (ASCR) and localized relative radiometric normalization (LRRN) techniques. Change/no-change thresholds for each method were optimized using calibration curves developed from reference data and a series of method-specific binary change masks. Cover class-specific thresholds were derived for each of the four methods using a previously developed NRB-LC classification (1998–1999) to support data stratification. An independent set of accuracy assessment points was selected using a disproportionate stratified sampling strategy to support the development of error matrices. Area-weighted conditional probability accuracy statistics were calculated based on the areal extent of change and no change for each cover class. All methods tested exhibited acceptable accuracies, ranging between 80% and 91%. However, change omission errors for woody cover types were unacceptably high, with values ranging between 60% and 79%. Overall commission errors in the change category were also high (42–51%) and strongly affected by the agriculture class. There were no significant differences in the Kappa coefficient between the NDVI, MID ASCR, and LRRN normalization methods. The MID non-normalized method was inferior to both the NDVI and MID ASCR methods. Stratification by major LC type had no effect on overall accuracies, regardless of method.