Structured dataset of reported cloud seeding activities in the United States (2000-2025) using an LLM

TL;DR

This paper introduces a comprehensive, high-accuracy dataset of U.S. cloud seeding activities from 2000 to 2025, created using an innovative PDF-to-text pipeline and large language models, enabling better analysis of weather modification efforts.

Contribution

The work presents a novel, scalable method combining LLMs and PDF extraction to compile a detailed cloud seeding dataset, filling a significant data gap in environmental research.

Findings

Dataset contains 832 records with 98.38% accuracy

Demonstrates LLMs' effectiveness in extracting structured data from PDFs

Provides a publicly available resource for climate and weather modification studies

Abstract

Cloud seeding, a weather modification technique used to increase precipitation, has been practiced in the western United States since the 1940s. However, comprehensive datasets are not currently available to analyze these efforts. To address this gap, we present a structured dataset of reported cloud seeding activities in the U.S. from 2000-2025, including the project name, year, season, state, operator, seeding agent, apparatus used for deployment, stated purpose, target area, control area, start date, and end date. Combining our multi-stage PDF-to-text extraction pipeline with OpenAI's o3 large language model (LLM), we processed 832 historical reports from the National Oceanic and Atmospheric Administration (NOAA). The resulting dataset demonstrates 98.38% estimated accuracy, based on manual review of 200 randomly sampled records, and is publicly available on Zenodo. This dataset addresses the gap in cloud seeding data and demonstrates the potential for LLMs to extract structured information from historical environmental documents. More broadly, this work provides a scalable framework for unlocking historical data from scanned documents across scientific domains.