Weather forecasting requires a comprehensive analysis of various types of meteorology data, and with the wide application of deep learning in various fields, deep learning has proved to have powerful feature extraction capabilities. In this paper, from the viewpoint of an image semantic segmentation problem, a deep learning framework based on semantic segmentation is proposed to nowcast Cloud-to-Ground and Intra-Cloud lightning simultaneously within an hour. First, a dataset with spatiotemporal features is constructed using radar echo reflectivity data and lightning observation data. More specifically, each sample in the dataset consists of the past half hour of observations. Then, a Light3DUnet is presented based on 3D U-Net. The three-dimensional structured network can extract spatiotemporal features, and the encoder–decoder structure and the skip connection can handle small targets and recover more details. Due to the sparsity of lightning observations, a weighted cross-loss function was used to evaluate network performance. Finally, Light3DUnet was trained using the dataset to predict Cloud-to-Ground and Intra-Cloud lightning in the next hour. We evaluated the prediction performance of the network using a real-world dataset from middle China. The results show that Light3DUnet has a good ability to nowcast IC and CG lightning. Meanwhile, due to the spatial position coupling of IC and CG on a two-dimensional plane, predictions from summing the probabilistic prediction matrices will be augmented to obtain accurate prediction results for total flashes.
10 October 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Synthetic aperture radar tomography (TomoSAR) can acquire high resolution in height direction by forming a large synthetic aperture along the tomographic direction. Compressed sensing (CS) is widely utilized in TomoSAR imaging to reduce the costs of data sensing. Nevertheless, traditional CS-based algorithms are limited to computational complexity and the non-trivial parameters' tunning. To address such problems, an efficient unfolded deep shrinkage-thresholding network (UDST-net) is proposed for TomoSAR imaging in this paper. The proposed method adopts convolutional neural network module to learn a generalized nonlinear sparse transformation operator, showing great benefits in exploring the optimal prior. Besides, the hyper-parameters of the optimization framework are learned by end-to-end learning mechanism instead of manual-defined which obviously improves the efficiency of imaging process. Inspired by residual network (ResNet), the residual learning is introduced to reconstruction blocks of the proposed imaging network, improving the robustness of the network. In addition, the training dataset is constructed from point cloud data based on TomoSAR imaging principles, enhancing the network's ability to extract structural information. Finally, extensive simulation and measured experimental results show the effectiveness of the proposed method, obtaining high-quality imaging results while maintaining high computational efficiency.
Most lightning location networks obtain the position results by optimizing the goodness of fit to determine that all combinatorial time differences of arrivals (TDOAs) are due to a common discharge. This paper proposes a three-dimensional (3D) lightning location method based on range difference (RD) space projection. The proposed method projects all the measurements into the RD space, which has the space-invariant feature of the equivalence cell and can be partitioned soundly. Aiming at the problem of computational cost of the procedure of the projection, the hierarchical strategy is proposed to improve computational efficiency. The performance of the RD space projection based on the hierarchical strategy is analyzed via Monte-Carlo simulations. The results show that the proposed method can locate lightning sources in real time with high accuracy. The results also show that the location accuracy is limited by the level of the inherent time uncertainty, the layout, and the size of the receiver network. Under the fixed layout and size of the receiver network, and the fixed measurement noise uncertainty, the positioning precision cannot be improved more even if the grid step is small enough or the number of receivers is large enough.
25 December 2024 in IEEE Geoscience and Remote Sensing Letters
With global warming driving more extreme weather events, real-time monitoring and forecasting of severe convective weather are increasingly critical. This paper presents a Multi-Task Generative Adversarial Network (MT-GAN) to improve nowcasting (short-term prediction) of severe weather, specifically precipitation and lightning. MT-GAN forecasts with 6-minute temporal resolution and up to 60-minute lead times. Its dual-stream architecture handles radar reflectivity and lightning data, using an encoder-decoder framework with SimVP for temporal feature extraction. A two-level feature fusion enhances spatio-temporal predictions, while a Patch-D discriminator improves realism. Tested on a central China dataset, MT-GAN outperforms ConvLSTM and PredRNN++ in image quality and nowcasting accuracy, with stronger emphasis on convective regions for better lightning initiation detection.
26 April 2025 in International Journal of Modern Physics C