A dataset of aligned RGB and multispectral UAV imagery for semantic segmentation of weedy rice

This article introduces a curated UAV dataset for detecting and segmenting weedy rice in cultivated fields. It includes 734 high-resolution RGB images along with their geospatially aligned multispectral (MS) counterparts, which feature four spectral bands: Green, Red, Red Edge, and Near-Infrared. The RGB images were annotated with polygon masks created by a fine-tuned Segment Anything model, then reviewed and corrected by experts. All images were resized to 1280 × 960 pixels and further processed to remove samples with missing or excessively extensive annotations (>90 %). The final dataset reflects various infestation levels, with weedy rice coverage ranging from under 5 % to nearly 90 %, supporting the development of robust models in diverse field conditions. All images were acquired in rice fields across Vietnam's Mekong Delta region during three consecutive cropping seasons using consumer-grade UAVs equipped with both RGB and MS sensors. This dataset provides extensive spectral and spatial information, making it a valuable resource for research in precision agriculture, including multimodal semantic segmentation, vegetation classification, and weed detection. It also facilitates benchmarking and domain adaptation studies to improve model generalization across different modalities.