Radiometric normalisation of multisensor/multitemporal satellite images with quality control for forest change detection

This paper aims to investigate the applicability of a relative radiometric normalisation method to a set of multitemporal images acquired by sensors of substantially different characteristics. The overall aim of the project is to assess the potential of satellite remote sensing for identifying forest land cover change in Scotland. In this study, the Pseudo-Invariant Features (PITS) concept was investigated. PITs are landscape elements (pixels) whose reflectance values are nearly constant over time. We use the Principal Components Analysis (PCA) to identify the PIFs, because of the simplicity of the approach and the accuracy of the results. The approach also needs less image interpretation, thus it saves time and offers objectivity in the selection of PIFs. The radiometric normalisation method of PIFs is applied on Landsat TM (1989 & 1994) and ETM+ (2000), IKONOS (2003) and DMC (Disaster Management Constellation) (2005) multitemporal images. The particular sensors are very diverse in spatial, spectral and radiometric information content. The images were radiance corrected and then orthorectified. Different orthorectification methods were used but the overall accuracy remained within change detection limits (+/- 0.5 pixels). The quality control of the radiometric normalisation is done spatially, spectrally and statistically. Issues about the use of PCA for identifying PIFs are discussed. The results show that the relative radiometric normalisation using PCA to select PIFs can perform very well in a multisensor/multitemporal application when care is taken in the pre-processing stages.