A Multimodal Image-Text Dataset for Potted Vegetables

Pot cultivation offers advantages such as environmental control and resource intensiveness, making it widely applicable in facility-based vegetable production. However, its key operational steps still heavily rely on manual labor, creating an urgent need for intelligent technologies to achieve precise management. Currently, most methods for recognizing the growth status of facility vegetables are based on single-modal RGB images, which are susceptible to lighting variations, plant occlusion, and substrate background interference in complex pot cultivation environments, making it difficult to be put into operation in real production scenarios. Moreover, existing agricultural datasets are mostly focused on open-field or orchard settings, lacking multimodal image-text data resources, specifically for potted vegetables. This limitation hinders the development and application of multimodal fusion technologies and vertical-domain large models in this field. To address the issues, this study constructs a multimodal image-text dataset of potted vegetables covering multiple crop types and growth stages, 185.36 GB in total. The image data includes three modalities: RGB, depth, and near-infrared, supporting several tasks, e.g., object detection, instance segmentation, and multimodal fusion, thereby meeting the development requirements of the visual modules within intelligent agricultural machinery. The text data comprises two categories: scene semantic descriptions and agricultural knowledge Q&A, which support the development and application of vertical-domain large models in potted vegetable scenarios. This dataset has been strictly registered and meticulously annotated, demonstrating excellent performance on the YOLOv11, GPT-4V, and LLaVA-NeXT models.