Deep Unfolding for Heterogeneous Change Detection: A Latent Calibration and Structured Separation Approach

Abstract

Heterogeneous change detection (HCD) is essential for Earth's surface monitoring with multisource remote sensing data, especially in scenarios where strictly homogeneous bitemporal observations are unavailable. However, the significant modality gap caused by different imaging mechanisms makes direct comparison unreliable, as unchanged regions may still exhibit pronounced cross-modal discrepancy. Existing methods have made progress in feature alignment and representation learning, yet they still face two persistent difficulties: establishing statistically calibrated and comparable cross-modal representations, and isolating true change responses from modality-induced interference during change extraction. To address these limitations, this article proposes an iterative collaborative decomposition and fusion network (ICDF-Net), which formulates HCD as a latent representation decomposition problem and integrates latent calibration with structured decomposition through a unified end-to-end framework. Specifically, an adaptive heterogeneous collaborative module is designed to construct more stable cross-modal latent representations by performing collaborative feature calibration. Based on the calibrated latent representations, a deep unfolding low-rank and sparse decomposition module is introduced to promote the separation of low-rank background and sparse residual components. The sparse residual components are then combined with differences between the calibrated representations and reconstructed by a hierarchical sparse reconstruction decoder to obtain the final change map. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed ICDF-Net, which achieves the highest F1-scores with low model complexity and competitive inference efficiency.