Correspondence analysis for detecting land cover change

The correspondence analysis (CA) method was applied to two multitemporal Landsat images of Raleigh, North Carolina for land use land cover (LULC) change detection. After the spectral transformation of the individual date images into component space using CA, the first component (PC1) of the date 1 image was subtracted from the PC1 of the date 2 image to produce difference image highlighting change areas. Accuracy curves based on the cumulative Producerʹs and Userʹs accuracies were then used to optimally locate threshold (cutoff) values in the high-end and low-end tails of the difference imageʹs histogram. Results were then compared to the standardized and non-standardized Principal Component Analysis (PCA) differencing and Normalized Difference Vegetation Index (NDVI) differencing methods for change detection. Results showed that there was 6.8% increase in urban related cover types in Raleigh metropolitan area between 1993 and 1999. Also, maps based on the CA differencing method were found to be thematically more accurate than maps based on PCA component differencing methods. Overall accuracy of change map produced by the CA method for the Raleigh metropolitan area was 92.5% with overall Kappa value of 0.88. In general, CA was found to be a powerful multivariate analysis technique when applied to change detection.