<p>Rainfall-triggered landslides rank among the world’s deadliest geohazards, inflicting thousands of fatalities and billions of dollars in infrastructure losses. Yet, current empirical and physically based models struggle to deliver large-scale operational warnings, partly because they reduce rainfall to cumulative or hypothetical patterns and ignore its intrinsic randomness. Here, we introduce a novel deep learning framework for landslide forecasting, rigorously benchmarked against logistic regression and validated with a deterministic slope stability model. A unique dataset spanning 1984–2022 was built, capturing triggering randomness and essential spatiotemporal features of natural slopes in Hong Kong. The framework integrates time-dependent modeling, tailored preprocessing to address real-world spatial and temporal imbalances, and a dual-threshold approach for nuanced risk gradation. Results indicate superior performance with an area under the ROC curve up to 0.92, boosting overall forecast accuracy by 36% over a random guess and increasing precision by 7% relative to rainfall-only predictors. Its probabilistic outputs closely align with deterministic analyses, yielding an operationally effective warning zone than that derived from a typical 0.5 cutoff. We package these advances into an interactive interface that ingests rain gauge readings, meteorological forecasts, and static terrain data for early warnings. This system paves the way for scalable, adaptive landslide risk management under increasingly erratic rainfall regimes.</p> Graphical Abstract <p></p>