A Long Time-Series Radiometric Normalization Method for Landsat Images

Radiometric normalization attempts to normalize the radiomimetic distortion caused by non-land surface-related factors, for example, different atmospheric conditions at image acquisition time and sensor factors, and to improve the radiometric consistency between remote sensing images. Using a remote sensing image and a reference image as a pair is a traditional method of performing radiometric normalization. However, when applied to the radiometric normalization of long time-series of images, this method has two deficiencies: first, different pseudo-invariant features (PIFs)—radiometric characteristics of which do not change with time—are extracted in different pairs of images; and second, when processing an image based on a reference, we can minimize the residual between them, but the residual between temporally adjacent images may induce steep increases and decreases, which may conceal the information contained in the time-series indicators, such as vegetative index. To overcome these two problems, we propose an optimization strategy for radiometric normalization of long time-series of remote sensing images. First, the time-series gray-scale values for a pixel in the near-infrared band are sorted in ascending order and segmented into different parts. Second, the outliers and inliers of the time-series observation are determined using a modified Inflexion Based Cloud Detection (IBCD) method. Third, the variation amplitudes of the PIFs are smaller than for vegetation but larger than for water, and accordingly the PIFs are identified. Last, a novel optimization strategy aimed at minimizing the correction residual between the image to be processed and the images processed previously is adopted to determine the radiometric normalization sequence. Time-series images from the Thematic Mapper onboard Landsat 5 for Hangzhou City are selected for the experiments, and the results suggest that our method can effectively eliminate the radiometric distortion and preserve the variation of vegetation in the time-series of images. Smoother time-series profiles of gray-scale values and uniform root mean square error distributions can be obtained compared with those of the traditional method, which indicates that our method can obtain better radiometric consistency and normalization performance.