Enhanced Temperature Nowcasting via Conv-LSTM Frame wise Modeling

This research introduces an enhanced method for temperature nowcasting through framewise modeling using a Convolutional Long Short-Term Memory (Conv-LSTM) architecture. Unlike traditional numerical models that often fall short in capturing the complex spatiotemporal dynamics of atmospheric data, the proposed approach leverages convolutional layers to extract spatial dependencies and LSTM units to learn temporal sequences, enabling precise short-term temperature prediction. The model is trained on sequential temperature frame data and evaluated using key performance metrics, achieving a Mean Squared Error (MSE) of 0.00035, Peak Signal-to-Noise Ratio (PSNR) of 34.54, Root Mean Square Error (RMSE) of 0.027, and Structural Similarity Index (SSIM) of 0.9954. These metrics demonstrate that the model delivers highly accurate predictions while maintaining the structural integrity and visual quality of the original temperature frames. The exceptionally high SSIM value highlights the model’s ability to preserve spatial consistency, which is vital in meteorological applications. This research underscores the potential of deep learning-based spatiotemporal modeling for accurate and reliable temperature nowcasting and offers a robust framework that can be further extended to other weather prediction tasks requiring fine-grained spatial-temporal understanding.