Regression functions for multitemporal relative calibration of Thematic Mapper data over boreal forest

Abstract

Different types of regression functions for multitemporal relative calibration of Landsat TM data were compared. The aim was to evaluate suitable function types for a proposed change detection technique. In this technique, multitemporal relatively calibrated satellite data are compared within spectrally homogeneous forest stands. In boreal forests, only minor parts of a large forest area are likely to substantially change within a few years. Thus, for change detection in forest stands, the forest land itself might be used as spectral reference for statistical image normalization techniques. The regression functions were used to predict the pixel values for the TM data at a recent acquisition from TM data obtained during an earlier acquisition. All functions were validated with a set of pixel values for forest without changes. No large error reductions were obtained with the use of ancillary forest data. Regression functions computed from only forest pixels gave higher coefficients of determination than functions computed from all image pixels or from dark and bright areas only. The relationship between spectral data from different years was approximately linear. Robust regression performed better than the least squares method when outliers were present. Also an iterated histogram matching was found to give an accuracy similar to that of band-to-band robust regression. Regression methods are sensitive to geometrical errors between spatially registered imagery, but histogram matching is not. Multiple regression functions from all TM bands were considerably better than the band-to-band regressions, especially for Bands TM 1–3. Thus when the geometric errors are small, multiple regression may be used instead of histogram matching. The good performance of the multiple regression also suggests that this method could be used even for change detection between imagery from different sensors.