Fractional vegetation cover in northern Saudi Arabian drylands: 1,155 quadrat photographs with trained-observer reference values and zero-shot multimodal model estimates

This dataset contains the field photographs, annotations, and model responses from a study evaluating whether zero-shot multimodal large language models can estimate fractional vegetation cover from a single ground-level quadrat photograph. Six open-weight models were each given four prompt designs and three image preparations across 1,155 quadrat photographs from arid rangelands in northern Saudi Arabia, with no fine-tuning and no worked examples, and compared against trained-observer estimates and a classical ExG−ExR color-index baseline. Included are the 1,155 photographs in all three variants, manual quadrat-corner annotations, scene-composition annotations for the material that most often confuses cover estimation, trained-observer reference values with per-image acquisition metadata, the four prompts verbatim, all 83,160 local and 27,720 API model estimates with parsed confidence values, a determinism re-run of a 100-image subsample, and the full text of every model response. Every table keys on the photograph filename, so any number in the paper can be traced back to the response and the image that produced it. See README.md for column definitions and for three details that affect reuse: the corner coordinates are given in the EXIF-oriented frame, the base archive holds original-resolution images rather than the downscaled copies sent to the models, and confidence is reported on a raw scale in the response logs and a normalized scale in the prediction tables.