Comparison of relative radiometric normalization methods using pseudo-invariant features for change detection studies in rural and urban landscapes

Relative radiometric normalization (RRN) to remove sensor effects, solar and atmospheric variation from at-sensor radiance values is often necessary for effective detection of temporal change. Traditionally, pseudo-invariant features (PIFs) are chosen subjectively, where as an analyst manually chooses known objects, often man-made, that should not change over time. An alternative method of selecting PIFs uses a principal component analysis (PCA) to select the PIFs. We compare the two RRN methods using PIFs in multiple Landsat images of urban and rural areas in Australia. An assessment of RRN quality was conducted including measurements of slope, root mean square error, and normalized difference vegetation index. We found that in urban areas both methods performed similarly well. However, in the rural area the automated PIF selection method using a PCA performed better due to the rarity of built features that are required for the manual PIF selection. We also found that differences in performance of the manual and automated methods were dependent on the accuracy assessment method tested. We conclude with a discussion on the relative merits of different RRN methods and practical advice on how to apply the automated PIF selection method