Multispectral Crop–Weed Segmentation Patches

This dataset provides a modified, patch-based version of the WeedsGalore dataset for crop–weed semantic segmentation using UAV multispectral imagery. The dataset consists of 224×224 image patches extracted from UAV acquisitions and includes RGB, near-infrared (NIR), and red-edge (RE) spectral bands, as well as five vegetation indices: NDVI, EVI, MSAVI, NDRE, and GNDVI. All spectral bands and indices are stored as single-channel grayscale images encoded as int32. Pixel-wise segmentation masks are provided as single-channel grayscale images (int8) with three classes: background (0), crop (1), and weed (2). To improve model robustness, data augmentation was applied to the training set only, including horizontal and vertical flipping, rotations of 90°, 180°, and 270°, and random zoom-in and zoom-out within ±20%. Validation and test samples were not augmented to prevent data leakage. This dataset is derived from the original WeedsGalore dataset (Celikkan et al., WACV 2025), which is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. The modified dataset is distributed under the same license and requires appropriate attribution to the original authors. The dataset is associated with the article:“Multispectral Feature Decomposition via Dual-Branch CNN–Transformer U-Net for Crop–Weed Semantic Segmentation”.