In recent decades, shifts in the spatiotemporal patterns of precipitation and extreme temperatures have contributed to more frequent droughts. These changes impact not only agricultural production but also food security, ecological systems, and social stability. Advanced techniques such as machine learning and deep learning models outperform traditional models by improving meteorological drought prediction. Specifically, this study proposes a novel model named the multivariate feature aggregation-based temporal convolutional network for meteorological drought spatiotemporal prediction (STAT-LSTM). The method consists of three parts: a feature aggregation module, which aggregates multivariate features to extract initial features; a self-attention-temporal convolutional network (SA-TCN), which extracts time series features and uses the self-attention module’s weighting mechanism to automatically capture global dependencies in the sequential data; and a long short-term memory network (LSTM), which captures long-term dependencies. The performance of the STAT-LSTM model was assessed and compared via performance indicators (i.e., MAE, RMSE, and R\(^2\)). The results indicated that STAT-LSTM provided the most accurate SPEI prediction (MAE = 0.474, RMSE = 0.63, and R\(^2\) = 0.613 for SPEI-3; MAE = 0.356, RMSE = 0.468, and R\(^2\) = 0.748 for SPEI-6; MAE = 0.284, RMSE = 0.437, and R\(^2\) = 0.813 for SPEI-9; and MAE = 0.182, RMSE = 0.267, and R\(^2\) = 0.934 for SPEI-12).

The increasing frequency of extreme weather events has resulted in a rapid growth of unstructured meteorological texts, posing challenges for accurate disaster-related information extraction. Existing relation extraction models that integrate semantic and syntactic information via Graph Convolutional Networks (GCNs) often suffer from noise amplification and information redundancy. To address these limitations, we propose a domain-specific relation extraction framework that integrates Bidirectional Encoder Representations from Transformers (BERT) with an Adaptive Denoising Graph Network. A dual-tuning architecture is designed to jointly model semantic and syntactic features. Specifically, a Self-Attention Graph Convolutional Network (SA-GCN) dynamically prunes irrelevant dependency structures, while a Residual Shrinkage Network (RS-Net) performs fine-grained feature denoising through sample-adaptive threshold learning. Experiments on the public DUIE 2.0 dataset and a newly constructed Agricultural Meteorological Disaster dataset demonstrate that the proposed model achieves F1 scores of 91.51% and 92.84%, respectively, outperforming state-of-the-art methods. Ablation studies further confirm the effectiveness of adaptive denoising in handling complex linguistic structures in disaster reports. The proposed approach enhances the robustness and accuracy of meteorological relation extraction and provides technical support for knowledge graph construction and intelligent disaster early warning systems.