Early Identification of Subtle Deformations in Potential Debris Flow Source Areas Using Phase-Unwrapped Convolutional Neural Networks and Long-Time-Series InSAR Technology

Mudslides are sudden and highly destructive; their source areas typically undergo slow, millimeter-scale creep over a period of months or even years before destabilization. If these precursor signals can be detected, valuable time can be gained for disaster prevention and mitigation. However, in the weathered crust and residual deposits of potential debris flow source areas, the long-term coupled action of freeze–thaw cycles and rainfall causes continuous reorganization of internal particle contact force chains, generating weak, metastable creep signals. The high-order nonlinearity and spatial heterogeneity of the interference phase gradient in low-coherence regions lead to pixel-spanning jumps in the unwrapped phase that are blurred by integer multiples of π. The high rate of phase jumps between adjacent pixels severely hampers the early detection of weak deformation. To address this, we propose a method for the early detection of weak deformation in potential debris flow source areas based on phase-unwrapping convolutional neural networks and long-time-series InSAR technology. First, we use long-time-series InSAR technology to construct a spatiotemporal map of interferogram sequences and establish feature propagation paths between high- and low-coherence interferogram pairs using the coherence coefficient as an edge weight. Second, we design a phase-unwrapping graph convolutional network that aggregates phase gradient information from neighboring nodes through two graph convolutional layers to correct the unwrapping results of low-coherence interferogram pairs and suppress cross-pixel jumps caused by π-integer-multiple blurring. Finally, by combining a dual-criterion classification approach based on temporal attention scores and deformation acceleration, the method captures the complete evolutionary process from stable creep to accelerated deformation. Experimental results show that the maximum phase jump rate of this method is approximately 0.02, effectively resolving the phase jump issue caused by high-order nonlinear gradients; in some areas of the study region, where deformation ranges from −1 mm to −9 mm, the inversion error is consistently controlled within ±1 mm. A total of five potential debris flow source areas were identified, classified by creep stage as follows: one in the accelerated deformation stage, two in the stable creep stage, and two in the early creep stage. No significant surface failure occurred in any of these source areas. This method provides reliable technical support and a decision-making basis for refined early warning, disaster prevention, and mitigation of debris flow hazards and holds significant engineering application value.