Data for: Research on the Scale Effects of UAV Flight Height and Crop Single Plant Mapping for Precision Agriculture

This study utilized a DJI Mavic 2 Pro drone to collect visible light remote sensing images of tobacco seedlings on June 6, 2023, from 14:00 to 16:00 under sunny weather conditions. The drone was equipped with a 20-megapixel Hasselblad camera and a 1-inch CMOS sensor, ensuring excellent imaging quality. The flight settings included a 80% overlap rate in both the forward and side directions, and the drone adopted waypoint hovering for photography to avoid motion blur and ensure clear images and geometric accuracy. Additionally, ground control points collected using RTK were integrated with Pix4D software for geographic registration and image stitching, generating digital orthophotomap images at heights ranging from 5 to 100m (every 5m), providing high-quality data support for the study.
This study employed high-resolution drone imagery and ArcMap 10.2 software for crop marking, significantly enhancing work efficiency compared to traditional field sampling methods. To address the issue of large drone data volume and limited computer processing capacity, tobacco imagery was first manually annotated and binarized, and then randomly cropped into 1024×1024 pixel samples to construct a precise tobacco dataset. To explore the impact of flight height variations on model recognition, imagery at four flight heights of 5m, 35m, 65m, and 95m was annotated to construct corresponding single-flight-height datasets. Simultaneously, to test the model's adaptability to different flight heights, the data from these four flight heights were merged into a multi-flight-height fused dataset, forming a total of five training datasets for the training and research of eight deep learning models: U-Net, U-Net++, LinkNet, PSPNet, MANet, FPN, PAN, and DeepLabV3+.