"BAWSeg: A UAV Multispectral Benchmark for Barley Weed Segmentation"

"BAWSeg is a large-scale UAV multispectral benchmark dataset for field-level barley weed semantic segmentation, constructed from repeated RTK-enabled UAV surveys over two commercial rainfed barley paddocks near Kondinin, Western Australia (E2: lat \u221232.508363, lon 118.338139; E8: lat \u221232.516563, lon 118.353799). Data were acquired across four growing seasons (2020\u20132023) using a DJI Phantom 4 Multispectral (P4M) platform with six synchronized cameras: five monochrome narrowband sensors (Blue\/Green\/Red\/RedEdge\/NIR) at 1600\u00d71300 pixels (focal length 5.74 mm, pixel pitch 2.74 \u03bcm) and one nadir RGB camera at 4000\u00d73000 pixels. All missions used nadir viewing (gimbal \u221290\u00b0), flight altitude 120 m AGL, 80% forward\/side overlap, and a commanded groundspeed of 5 m\u00b7s\u207b\u00b9. A D-RTK 2 base station provided real-time corrections, and per-frame exposure-time geotags were written into image EXIF via TimeSync. Each sortie includes radiometric reference panel captures (Spectralon, 1 m\u00d71 m) and per-exposure irradiance measurements to support repeatable conversion to surface reflectance.This DataPort release contains (i) raw per-frame multispectral and RGB imagery with complete acquisition metadata, (ii) mission boundary and flight-line geometry, (iii) radiometrically calibrated, band-coregistered orthomosaics exported as tiled GeoTIFF reflectance maps for all five bands, (iv) orthomosaic-level vegetation index rasters (NDVI, GNDVI, EVI, SAVI, MSAVI), (v) dense pixel-level semantic annotations for crop\/weed\/other with an explicit ignore label for out-of-field\/invalid pixels, and (vi) a leakage-free benchmark protocol based on non-overlapping 1024\u00d71024 spatial blocks and a fixed block split reused across years, plus training-ready 256\u00d7256 patches (stride 128) with per-patch metadata. BAWSeg enables reproducible evaluation under within-plot, cross-plot, and cross-year settings for precision weed mapping and robust multispectral segmentation research. The total data volume exceeds 500 GB, preserving full traceability from raw frames to benchmark-ready patches."