A visible light-thermal infrared multimodal image dataset collected from a real greenhouse tomato cultivation scene for flower and fruit detection

The rapid development of machine vision has promoted the automation of crop monitoring, providing technical support for planting management. For this purpose, this paper constructs a visible light-thermal infrared multimodal image dataset for the detection of flowers and fruits of facility tomatoes: it includes 300 images of the flowering period and 343 images of the fruiting period collected by the H982 visible light camera, as well as 145 images of the fruiting period collected by the HIKMIC thermal infrared camera. The LabelImg tool is used to complete the bounding box annotation of tomato flowers, fruits, and fruit maturity. Based on the evaluation of YOLOv11 (n/s/l) and YOLOv13 (n/s/l), the optimal mAP50 for the flower subset can reach 89.86%, while the optimal mAP50 for the fruit subset and the entire dataset is above 84%. The overall detection accuracy is excellent, indicating that the dataset is practical. The multimodal fusion experiment shows that compared with single visible light, the fusion strategy significantly improves the detection performance without sacrificing the inference efficiency almost at all, verifying the complementarity of thermal infrared and visible light, and providing data support for the research and deployment of intelligent monitoring algorithms for facility tomatoes.