RealWheat2021

The dataset is approximately 230 GB in size, compressed into 12 volumes and stored across 6 Zenodo repositories. A real wheat dataset (RealWheat2021) was constructed for network development and performance evaluation. First, the original UAV images were stitched and geometrically calibrated with the 11 GCPs using the Agisoft Metashape Professional 1.5.1 software (Agisoft LLC, St. Petersburg, Russia). The orthophotos were manually adjusted to the north-south direction, and regions of interest were then clipped using the ArcMap 10.8 software (ESRI, Redlands, USA). A total of ten orthophotos belonging to six growth stages were generated. Second, the ten orthophotos were partitioned into training and testing datasets, respectively, adhering to a ratio of 7:3. Notebally, the testing datasets include three Stages for studying the robustness of proposed methods on different growth stages. Third, the ArcGIS software was employed to label each plot within the orthophotos. Fourth, each orthophoto and its label file were cropped into patches using the ArcMap software. A predefined sliding window approach was employed to extract patches from each orthophoto and its corresponding label file, with each patch having a consistent width and height of 1024 pixels, encompassing approximately four plots. The sliding step of the window was set as 0 to avoid any intersection of each patch. Fifth, the patches were rotated for data augmentation. The patches cropped from the orthophotos were defined as the initial direction (0°). Then, each patch was rotated clockwise starting from 0° to 90° with an interval of 15° (0°, 15°, 30°, 45°, 60°, 75°, 90°) using a Python package named OpenCV. The corresponding labels were processed in the same way as the patches. Finally, the RealWheat2021 dataset was constructed with 115234 images.本数据集总容量约230GB，压缩为12个分卷，存储于6个Zenodo仓储中。 为开展网络开发与性能评估，本研究构建了真实小麦数据集（RealWheat2021）。首先，采用俄罗斯圣彼得堡Agisoft LLC公司的Agisoft Metashape Professional 1.5.1软件，结合11个地面控制点（Ground Control Point，GCP）对原始无人机影像进行拼接与几何校正；随后将正射影像手动调整为南北朝向，并使用美国雷德兰兹的ESRI公司的ArcMap 10.8软件裁剪感兴趣区域，最终生成覆盖6个生育期的共10张正射影像。其次，按照7:3的比例将10张正射影像划分为训练集与测试集，值得注意的是，测试集包含3个生育期，用于研究所提方法在不同生育期下的鲁棒性。第三，借助ArcGIS软件对正射影像中的每个样区进行标注。第四，使用ArcMap软件将每张正射影像及其对应的标签文件裁剪为图像块：本研究采用预设的滑动窗口方法从正射影像及其对应标签文件中提取图像块，每个图像块的宽高均为1024像素，涵盖约4个样区；为避免图像块间出现重叠，将滑动步长设为0。第五，对图像块进行旋转以实现数据增强：将从正射影像中裁剪得到的图像块初始方向定义为0°，随后使用名为OpenCV的Python工具包，将每个图像块以15°为间隔从0°顺时针旋转至90°（即0°、15°、30°、45°、60°、75°、90°），对应的标签文件采用与图像块完全一致的处理方式。最后，RealWheat2021数据集共包含115234张图像。