BiDepth: A Bidirectional-Depth Neural Network for Spatio-Temporal Prediction

Accurate spatio-temporal (ST) forecasting is critical for dynamic systems such as urban mobility and weather, yet existing models struggle to capture non-stationary correlations while remaining computationally efficient. To address this, we propose the BiDepth Multimodal Neural Network (BDMNN), which integrates two novel mechanisms: (1) a bidirectional depth modulation strategy that dynamically allocates computational complexity to balance long-term seasonal trends and short-term fluctuations; and (2) a Convolutional Self-Attention Cell (CSAC) that preserves spatial topology during temporal reasoning. On the real-world NYC Taxi and US Precipitation datasets, BDMNN outperforms standard baselines, achieving MSE reductions of up to 14.6%. Crucially, it matches or surpasses massive state-of-the-art video-prediction architectures (e.g., SwinLSTM, SimVP) while using as few as 0.44M parameters, versus up to 18M for the baselines. These results establish BDMNN as an efficient and robust solution for diverse real-world ST applications.