Automated Chain-of-Thought Data Generation for Meteorological Reasoning: A Study Based on WeatherQA

High-quality Chain-of-Thought (CoT) data plays a critical role in enhancing model reasoning in vertical domains. In meteorology, constructing such data is valuable yet difficult due to the complexity of visual data and the reliance on expert annotations. To address this, we propose an automated CoT data generation method based on a three-stage modality bridging pipeline: image description generation, pseudo-CoT generation, and final reasoning chain construction. Using the WeatherQA dataset, we build CoT data under single-image and three-image configurations and fine-tune Qwen2.5-VL-7B on two key tasks. Experimental results show that our CoT-trained model achieves comparable performance to expert-annotated baselines (such as 64.21% vs. 64.29% accuracy in Area Affected accuracy), demonstrating the effectiveness of automated CoT construction for meteorological reasoning.