Cyclone intensity estimation presents a complex challenge due to cyclonic systems’ dynamic and intricate nature. Traditional models often struggle with capturing the nuanced temporal variations of cyclone behavior from infrared imagery. To address these limitations, this paper introduces a novel approach utilizing a Dual-Stage Temporal Attention Mechanism (DSTAM) integrated with Recurrent Neural Networks (RNNs). DSTAM enhances the RNN’s capacity to discern and prioritize critical features by implementing a two-tiered attention mechanism: Context-Aware Attention and Adaptive Memory-Based Attention. The Context-Aware Attention stage computes a context vector to evaluate the relevance of each time step. At the same time, the Adaptive Memory-Based Attention adjusts focus based on the RNN’s evolving hidden states. This dual mechanism allows for a dynamic adjustment of attention, improving the model’s sensitivity to both overarching patterns and temporal changes in cyclone intensity. The proposed method was validated using a comprehensive dataset of infrared images sourced from the MOSDAC server, covering cyclone activity from 2012 to 2021. Empirical results demonstrate that DSTAM significantly enhances prediction accuracy compared to conventional models. The integration of DSTAM with RNNs yields superior performance in capturing cyclone intensity dynamics, offering a robust tool for meteorological forecasting. Future research directions include extending DSTAM to incorporate additional data modalities and exploring hybrid machine-learning frameworks for further improvements in cyclone intensity prediction.