<p>The occurrence of landslides is closely related to rainfall. Therefore, rainfall characteristics have long been recognized as a key influencing factor in landslide susceptibility assessments (LSA). However, the kilometer-level resolution rainfall data relied upon by traditional evaluation methods struggles to accurately capture local hydrological anomalies in complex terrain, resulting in limited spatial representation capabilities of the models. As a composite signal, InSAR coherence exhibits temporal variations primarily influenced by combined factors such as rainfall and vegetation. Therefore, this study constructs an InSAR coherence separation model to quantitatively extract the rainfall-dominated component from the coherence time series. This serves as a proxy feature for high-resolution hydrological disturbances, supporting more detailed LSA at finer scales. To systematically validate the feasibility and effectiveness of this method, the study constructed a long-term memory-gene model (LGM) and performed landslide susceptibility modeling (LSM) and comparisons with several other models. Experimental results indicate that within the LGM model framework, the prediction performance of high-resolution rainfall impact features derived from InSAR coherence separation is optimal, achieving an AUC value of 0.97. This significantly outperforms methods using raw coherence (AUC = 0.95) and kilometer-scale rainfall data (AUC = 0.94). Furthermore, the model demonstrated outstanding performance across comprehensive metrics, achieving accuracy, recall, and F1 scores of 0.90, 0.92, and 0.90, respectively. This research has fully demonstrated the superiority of the proposed method in LSA, not only significantly enhancing model performance but also providing a new technical approach to overcoming the limitations of rainfall data resolution.</p>