<p>A natural disaster with abrupt and intense downpours in particular areas lead to cloudburst. It may cause flash floods and landslides, resulting in huge lives and assets loss. Several methods such as Doppler radar, weather prediction models, satellite imagery and weather radar are utilized in cloudburst prediction. The complexity of cloudburst prediction is due to the restrictions of current infrastructures, such as real-time and high spatial-resolution data unavailability. By integrating Convolutional Neural Networks (CNNs) with Gramian Angular Field (GAF) encoding, a novel forecasting framework for Himalayan cloudbursts in Uttarakhand, India is presented in this study. The complex multivariate 1D time-series data is translated into 2D visual representations by the model using historical data from the Indian Meteorological Department. CNNs potential to extract spatial features is employed by the framework by converting this into a Multivariate Time Series Classification (MTSC) problem, efficiently detecting meteorological precursors essential for timely early warnings. For 6-h and 12-h lead durations with AUC of 97.73% and 98.41%, respectively, the model exhibits robust predictive ability when validated on historical data. Also, the issue of data scarcity inherent to these rare extreme events is explicitly addressed in this study by adopting CNN-GAF for MTSC data and leveraging rigorous cross-validation to provide statistical reliability despite limited ground-truth samples. In vulnerable mountainous terrains, it gives a scalable and computationally effective solution for disaster management. This study bridges the gap between operational disaster response and theoretical modelling by exhibiting the ability to extend early warning infrastructures with a real-time, lightweight, edge-deployable system that may be combined with low-resource IoT weather stations.</p>