Accurate short-term precipitation forecast is extremely important for urban flood warning and natural disaster prevention. In this paper, we present an innovative deep learning model named ISA-PredRNN (improved self-attention PredRNN) for precipitation nowcasting based on radar echoes on the basis of the advanced PredRNN-V2. We introduce the self-attention mechanism and the long-term memory state into the model and design a new set of gating mechanisms. To better capture different intensities of precipitation, the loss function with weights was designed. We further train the model using a combination of reverse scheduled sampling and scheduled sampling to learn the long-term dynamics from the radar echo sequences. Experimental results show that the new model (ISA-PredRNN) can effectively extract the spatiotemporal features of radar echo maps and obtain radar echo prediction results with a small gap from the ground truths. From the comparison with the other six models, the new ISA-PredRNN model has the most accurate prediction results with a critical success index (CSI) of 0.7001, 0.5812 and 0.3052 under the radar echo thresholds of 10 dBZ, 20 dBZ and 30 dBZ, respectively.

To address the current challenges in fire detection algorithms, including insufficient feature extraction, high computational complexity, limited deployment on resource-constrained devices, missed detections, false detections, and low accuracy, we developed a high-precision algorithm named GSF-YOLOv8. The GSF-YOLOv8 consists of three main parts: Firstly, we introduced the Gather-Distribute (GD) feature aggregation module, which enhances the recognition accuracy of fire features by fusing and extracting multi-scale information. Secondly, we integrated the SimAM attention module to optimize the collaborative learning of feature map dimensions and to reduce reliance on hyperparameter tuning through precise weight allocation, thereby improving the model’s rapid response capability to fires. Lastly, we proposed a Focal-DIoU Loss to replace the original loss function, optimizing bounding box regression and improving localization accuracy. It also focuses on hard samples to reduce the impact of easily misclassified samples, thereby enhancing the model’s performance in complex fire detection. We tested our method on the D-Fire dataset. The experimental results showed that GSF-YOLOv8 achieved higher mAP (80.65%) and faster detection speed (118 FPS). Meanwhile, the model’s complexity and number of parameters have been significantly optimized. The GSF-YOLOv8 not only significantly improves the efficiency and accuracy of fire detection but also provides a more reliable and accurate solution for real-time detection in similarly complex environments.

We propose the Enhanced Predictive Recurrent Neural Network (EnPredRNN) based on PredRNN to extend the period of spatio-temporal prediction. To better capture global spatial dependencies, we integrate the Self-Attention (SA) module into PredRNN’s basic unit, forming the Enhanced Long Short-Term Memory (EnLSTM). Aiming at the problem that the standard temporal memory state in PredRNN contains insufficient information about inter-frame motion, we propose the Enhanced Temporal Memory (ETM) by aggregating past multi-step temporal memory states. Aiming at the gradient vanishing problem in Recurrent Neural Networks (RNN), the Alleviating Gradient Vanishing (AGV) structure is used to construct the high-speed path that facilitates gradient propagation. Experimental results show that EnPredRNN effectively extends spatio-temporal prediction from ten to thirty time-steps.