Landslides triggered by rainfall are a significant hazard in mountainous regions worldwide, posing risks to both infrastructure and human safety. Projections regarding climate change indicate an increase in both extreme weather events and subsequent landslide incidents. Therefore, accurately forecasting these rainfall-induced landslides is essential for implementing effective hazard mitigation and evacuation strategies. Traditionally, predictions have relied on physically based models and empirical rainfall thresholds that account for both rainfall intensity and duration. With the introduction of Machine Learning, the ability to incorporate static factors such as slope gradients and soil classifications has been significantly improved, thereby enhancing predictive accuracy and enabling broader spatial applications. Nonetheless, recent research involving Machine Learning has predominantly concentrated on established deep learning frameworks, while innovative approaches have not been thoroughly investigated. This hesitance to embrace contemporary deep learning techniques may stem from challenges in interpreting the decisions made by these models, which are vital for effective operational landslide early warning systems. Recent studies emphasising traditional Machine Learning frequently include analyses of network behaviour through post-hoc interpretations, utilising methods such as Shapley values and assessments of feature importance. However, the use of inherently explainable deep learning networks for rainfall-induced landslide prediction remains underexplored. To address this gap, we propose employing Kolmogorov-Arnold Networks (KANs) to predict rainfall-induced landslides, leveraging precipitation time series obtained from a globally accessible satellite product. The proposed model achieves competitive performance compared to various established models while maintaining interpretability. In addition to utilising interpretable activation functions, we also suggest implementing Dynamic Rainfall Thresholds (DRT) as a visual interpretation tool for the model. This combination of interpretative tools, paired with a low rate of missed alarms, positions the model as a suitable option for critical applications such as landslide early warning systems.