Applying Relative Radiometric Normalization Methods To Multi-temporal And Cross-sensor Remote Sensing Images

Relative radiometric normalization of multi-temporal remotely sensed images is one of the basic procedures for remotely sensed data preprocessing.Currently,the discrepancies in solar illumination geometry,atmospheric condition and sensor performance frequently exist when acquiring multi-temporal satellite remote sensing images,which may lead to the change in radiation values at the same location over time.However,this change may not reflect the real physical alternation of the corresponding ground objects.Thus,it is necessary to implement relative radiation normalization processing for these multi-temporal remote sensing images before applying them to conduct land cover change detection or image mosaicking operations.The major objective of relative radiometric normalization among multi-temporal,cross-sensor images is to retain the radiation consistency of these images,or to make these images comparable directly.In this study,bi-temporal Landsat5 TM imagery acquired on August 19 th,2010 and Landsat8 OLI imagery acquired on October 14 th,2013,with path/row number of 120/038,covering Nanjing city were involved.First,the 2013 imagery was used as the reference to normalize the 2010 imagery by using pixel-based normalization methods including the pseudo invariant features(PIF),the temporally invariant cluster(TIC),the wall to wall regression(WTW),the multivariate alternation detection(MAD),the dark bright set normalization(DB),the automatic scattergram controlled regression(ASCR)and the random forest normalization methods(RF),followed by use of the distribution-based maximum minimum normalization method(MM),the mean standard normalization method(MS),the histogram matching(HM)and the ordinal conversion methods(OC).Then,their normalization performances of different relative radiation normalization methods were evaluated by using eight objective evaluation measures including the information entropy,the spectral distortion,the edge strength,the edge strength,the spatial frequency,the cross-entropy,the peak signal-to-noise ratio and the mutual information.Next,based on the normalized images generated from the optimal four relative radiation normalization methods,the land cover change detection was implemented by using the change vector analysis process,and the change detection accuracy was validated by visually interpreting those corresponding years' high spatial resolution Google Earth maps.Results showed that:(1)From the visual effect,the spectral characteristics of the normalized images were close to those of the reference image,and the spatial information of the normalized images remained intact without damaging the spectral characteristics of the ground objects.(2)After comprehensively comparing the objective evaluation index values,random forest based normalization method was the best,followed by the multivariate change detection,the temporally invariant cluster,the ordinal conversion method,the wall-to-wall regression,the pseudo-invariant features,the mean standard normalization method,the automatic scattergram-controlled regression,the histogram matching,the maximum minimum normalization method and the dark set-bright set normalization method.(3)For the change detection analysis,the overall accuracy of the obtained change detection results was ranked as follows: random forest normalization method(79%)greater than multivariate alternation detection method(76%),followed by temporally invariant cluster(74.33%),and mean standard normalization method(70.33%).This study concludes that the random forest based normalization method is more effective in radiometric normalization processing and change detection application,which can provide technical reference for engineering normalization of cross-sensor and multi-temporal remote sensing images.