An improved CNN-BiLSTM video-based rainfall intensity estimation method based on attention mechanisms and residual connection variants

Accurate monitoring of rainfall intensity is essential for forecasting and managing floods. The surveillance video-based rainfall measurement method has potential for high spatial and temporal resolution rainfall monitoring. Existing video-based rainfall estimation models have incorporated temporal information; however, challenges remain in improving accuracy and robustness under different environmental conditions. To address this issue, this paper proposes an improved hybrid convolutional neural network-attention-bidirectional long short-term memory (CNN-Attention-BiLSTM) model. The temporal feature extraction is achieved by stacked BiLSTM layers with a learnable fusion strategy and a residual connection employing pre-layer normalization. An efficient channel attention mechanism is integrated for critical information. Results show that the proposed model can achieve high accuracy and efficiency in rainfall intensity estimation under various conditions. Generalization capability and robustness are exhibited in well-illuminated daytime scenarios. This study holds significant implications for the establishment of video-based, high spatio-temporal resolution rainfall monitoring systems.