Hydrological report generation using multimodal models

Abstract

This thesis presents HydroGen, a domain-adaptive multimodal framework for hydrological report generation using satellite images and air-pressure maps. Existing multimodal large language models still face limitations in hydrological domain adaptation, sequential image understanding, geographic grounding, and cause–effect reasoning. To address these limitations, this study constructs HydroMM-Instruct from 431 weekly hydrological reports and 5,590 images collected from Thailand’s Hydro-Informatics Institute between 2018 and 2025. HydroGen integrates image preprocessing, two-stage hydrological training, Temporal Prompt Tokens, and GuideAC to improve multimodal report generation. The framework is evaluated using four rolling year-level cross-validation folds and compared with open-source and proprietary multimodal baselines. HydroGen achieves BERT-F1, Sentence Similarity, BLEU, and ROUGE-L scores of 84.56%, 79.52%, 62.01%, and 67.78%, respectively, outperforming all baseline models. Qualitative and expert evaluations further confirm that HydroGen improves factual correctness, causal reasoning, geographic consistency, temporal consistency, and completeness. These results show that domain adaptation, temporal prompting, and knowledge-guided reasoning are effective for improving multimodal hydrological report generation.