Traditional rainfall prediction models have limitations in adapting to seasonal precipitation variations and suffer from significant accuracy degradation under small-sample conditions in Xinjiang's arid regions. To address this, a prediction method integrating dynamic threshold optimization and multi-feature LightGBM was proposed. A 9-dimensional feature system was constructed, including precipitable water vapor (PWV) baseline values, PWV increments, PWV growth rates, and other auxiliary atmospheric variables. A monthly dynamic threshold modeling mechanism was adopted to address the dual challenges of seasonal rainfall differences and insufficient training samples. The results show that the dynamic threshold model achieves a critical success index of 85.5% in April, which is 7.5 percentage points higher than the annual optimal fixed threshold. The model maintains a detection rate above 92% for both convective rainfall in July and stratiform rainfall in September, with a false alarm rate below 22%. The effectiveness of the dynamic threshold mechanism and multi-feature fusion strategy is confirmed in adapting to seasonal rainfall patterns and improving prediction accuracy under limited sample conditions, thereby providing technical support for the prevention of highway geological disasters in the arid regions of Xinjiang.