MeteoSaver LLM based software for the transcription of historical weather data (egusphere.copernicus.org)

MeteoSaver v1.0 is an open-source, machine-learning driven tool designed to transcribe handwritten tabular weather observations into digital datasets. The software runs a five-stage pipeline—image pre-processing, table and cell detection, transcription, quality assessment/quality control, and final data formatting/upload—using user-defined configuration to adapt to different sheet layouts. In a validation on ten historical temperature sheets from the Democratic Republic of the Congo, MeteoSaver successfully transcribed 95–100% of records; a median 74.4% of transcriptions received the highest internal quality flag, 74% matched manually transcribed records, and the median mean absolute error was 0.3 °C. This matters for the AI/ML and climate communities because millions of archival meteorological sheets remain locked in paper form, especially in data-scarce regions. By combining layout-aware table detection with ML-based handwriting recognition and embedded QA/QC, MeteoSaver can scale digitization efforts, preserve long historical climate series, and accelerate research into regional climate trends. The tool’s open-source release (MeteoSaver v1.0) makes it extensible for different handwriting styles, table dimensions, and maintenance conditions, with best results when sheet formats are consistent—positioning it as a practical bridge between archival records and modern climate analysis.