白鹤滩水电站作为国家重大工程,其库区地质条件复杂,蓄水后岸坡稳定性监测至关重要。针对传统InSAR技术依赖形变速率指标,难以揭示时序演变规律的问题,本文融合迁移学习与时间卷积网络(TCN),构建了一种时序形变模式智能识别模型,旨在提升库区形变演化特征的认知。研究利用2021年4月至2024年3月的Sentinel-1卫星数据,采用SBAS-InSAR技术反演获取整个库区的时序形变场,提出了一种“迁移学习增强的深度学习时序分类架构”：利用模拟样本数据进行预训练,再通过真实样本数据进行微调,研究建立了一个包含平稳、线性、阶跃、分段线性、幂次及未定义类的六维形变分类体系,并应用于白鹤滩库区的形变模式识别。结果表明：(1)库区97,377个形变点分类统计显示：平稳型(45.6%)、分段线性型(24.4%)为主导模式,平稳型与低速率、高相干性显著相关；(2)石门坎滑坡受库水位滞时调控,抬升滞后蓄水峰值约2个月,野猪塘坡体呈现降雨触发的阶跃形变及水位–降水耦合的分段线性响应。本文突破了传统基于速率的单一分析框架,验证了迁移学习在降低时序标注成本方面的有效性,并表明多维形变模式解析能够为库区灾害风险的动态评估提供新的方法范式。

Objective: The Baihetan Hydropower Station, as a national mega-project, is located in a geologically complex reservoir area where slope stability monitoring after impoundment is crucial. Traditional InSAR techniques mainly rely on deformation rate indicators and thus face limitations in capturing the full temporal evolution of slope deformation. This study aims to overcome these limitations by developing an intelligent framework for time-series deformation pattern recognition to improve hazard identification and risk assessment in the reservoir area. Method: Sentinel-1 SAR data spanning April 2021 to March 2024 were processed using the SBAS-InSAR technique to retrieve the reservoir-wide deformation field. A transfer learning–enhanced deep learning architecture was proposed, in which synthetic samples were used for pretraining and real samples for fine-tuning. Based on this framework, a six-class taxonomy of deformation patterns—stable, linear, step-like, piecewise linear, power-law, and undefined—was established. A Temporal Convolutional Network (TCN) was then applied to classify deformation modes across the Baihetan reservoir. Results: (1) Among 97,377 deformation points, the dominant patterns were stable (45.6%) and piecewise linear (24.4%), with stable deformation strongly associated with low velocity and high coherence. (2) The Shimenkan landslide exhibited a delayed uplift response regulated by reservoir water level, lagging peak impoundment by about two months.(3) The Yezhutang slope showed rainfall-triggered step-like deformation and a piecewise linear response under coupled water level–precipitation forcing. Conclusion: The findings demonstrate that transfer learning effectively reduces labeling costs, while multidimensional deformation pattern analysis provides a new paradigm for dynamic risk assessment of reservoir-induced geohazards.