Unwrapping Is Not All You Need for Capturing the Temporal Evolution of Creeping Landslides

For deformation monitoring of creeping landslides, MT-InSAR is often limited by phase unwrapping, atmospheric correction, and spatiotemporal filtering, leading to complex workflows and elevated costs. This study proposes a wrapped-interferogram-driven framework for creeping landslide hazard identification and temporal evolution reconstruction. For single-scene recognition, we build the Jinsha Dataset and introduce an Adaptive Frequency-Domain Decoupling and Fusion Block (AFDFB) into the shallow encoder of a segmentation network to mitigate frequency mixing. By coupling convolution with State Space Model (SSM), AFDFB enhances the characterization of high-frequency fringe boundaries and low-frequency structural trends. For temporal modeling, we migrate time-series concepts from the phase domain to the pixel-wise probability domain and integrate coherence weighting, cyclic-consistency constraints, and temporal accumulation to produce an unwrapping-free quasi-time-series risk representation. The proposed network achieves 80.79% IoU and 89.37% F1-score on the test set. In both single-landslide assessment and regional inventory mapping, probability-domain fusion suppresses isolated anomalous responses and yields more stable risk maps, with spatial delineation and temporal variation trends consistent with MT-InSAR.