A multi-stage cascaded framework for relative radiometric normalization of high-resolution remote sensing imagery in karst regions

High-resolution multi-temporal remote sensing imagery is essential for land-cover mapping and long-term environmental monitoring, but radiometric inconsistencies often reduce the reliability of cross-date comparison and change analysis. This problem is particularly severe in karst regions, where rugged terrain, terrain-induced shadowing, seasonal phenological variation, rock – vegetation spectral coupling, and atmospheric heterogeneity make pseudo-invariant feature (PIF) extraction difficult and weaken conventional relative radiometric normalization (RRN) methods. To address these challenges, we propose a multi-stage cascaded RRN framework, termed GCCM, which integrates Gaussian mixture model (GMM)-based probabilistic screening, dual-branch convolutional neural network (CNN)-based deep feature learning, and canonical correlation analysis (CCA) – multivariate alteration detection (MAD)-based statistical refinement. GCCM first screens candidate invariant regions, then learns robust spectral – spatial representations for reference and target images, and finally refines invariant-region selection in the learned feature space. Experiments were conducted on multi-temporal GF-1 and Sentinel-2 image pairs over a karst area in northern Guangxi, China, under large-temporal-span and pronounced seasonal-variation conditions. The results show that GCCM generally outperforms SAD-ED and IR-MAD in radiometric normalization accuracy and PIF quality, although improvements vary among datasets and sensors. In the challenging GF-1 large-temporal-span experiments, GCCM achieved a mean R2 of 0.800 and a mean RMSE of 0.007. In the seasonal-variation and Sentinel-2 experiments, most band-wise R2 values exceeded 0.900 with low RMSE values. Ablation analysis further showed that the complete GCCM achieved a mean R2 of 0.9370 and an RMSE of 0.0126 on four representative datasets. These results indicate that GCCM provides a robust and terrain-adaptive solution for RRN in complex karst environments.