Mustard Flea Beetle Damage Dataset

Flea beetles (Phyllotreta spp.) are a major early-season pest of mustard (Brassica juncea) in Bangladesh, capable of causing economically significant defoliation within one to two days of onset. Images were captured across three mustard fields in two districts (Gazipur and Tangail), at varying times of day and planting densities ranging from sparse stands to frames containing over 200 overlapping seedlings, using four smartphone models (iPhone 16 Pro Max; Samsung Galaxy A34, M11, M32) to introduce natural variation in resolution, colour balance, and sharpness. The dataset has two parts: Detection has 1,603 field photographs, each manually annotated in Roboflow with a bounding box around every visible mustard seedling under a single class ("mustard plant"). Annotations follow the YOLO format: one .txt per image, each line giving class_index x_center y_center width height, normalised to [0, 1]. Only one class is present (index 0). Images are split 80:10:10 at the image level, so no plant appears in more than one split. Classification has 35,948 plant-level crops, extracted by applying the trained detector to all field images, padding each bounding box by 30 pixels, and resizing to a fixed square canvas with letterbox padding. Crops were manually sorted into four severity classes by the proportion of leaf area showing feeding damage: healthy (0–10%), mild (10–50%), moderate (50–70%), severe (70–100%). Contents: mustard_flea_beetle_dataset/ |── detection/ | |── images/ | | ├── train/ (1,282) | | ├── val/ (160) | | └── test/ (161) | |── labels/ | ├── train/ | ├── val/ | └── test/ |── classification/ ├── healthy/ (32,432) ├── mild/ (1,219) ├── moderate/ (927) └── severe/ (1,370) The class distribution reflects natural incidence in the sampled fields and is heavily skewed. The four-class severity scale contains an unlearnable boundary. Merging the moderate and severe classes substantially improved classification performance. How to interpret and use the data: The four original labels are preserved so users can apply their own severity scheme. Anyone reproducing the three-class formulation should merge moderate into severe. Several characteristics affect interpretation. Severity boundaries were assigned by visual estimation rather than measured leaf area. Annotation was performed by a single annotator, so inter-annotator agreement was not assessed. Classification crops carry black letterbox borders from aspect-ratio-preserving resizing. Generalisation to other regions, seasons, and mustard varieties is untested. All images are unmodified field photographs or crops derived from them. No synthetic or augmented images are included.跳甲属（*Phyllotreta* spp.）昆虫是孟加拉国芥菜型油菜（*Brassica juncea*）的主要季初害虫，其种群爆发后1至2天内即可造成具有经济损失规模的落叶危害。本次数据集的图像采集自加济布尔（Gazipur）与坦盖尔（Tangail）两个地区的三块芥菜田，拍摄时段覆盖全天不同时刻，种植密度跨度较大，从稀疏植株到单种植框内超过200株重叠幼苗不等；实验采用四款智能手机机型（iPhone 16 Pro Max；三星Galaxy A34、M11、M32）进行拍摄，以引入分辨率、色彩平衡与锐度的自然差异。 本数据集包含两个子任务模块： 1. 检测任务数据集：包含1,603张田间实拍照片，所有图像均在Roboflow平台上完成人工标注，为每个可见的芥菜植株绘制边界框，仅设置单个类别“芥菜植株”。标注格式遵循YOLO（You Only Look Once）规范：每张图像对应一个.txt格式标注文件，文件内每行标注格式为"class_index x_center y_center width height"，所有坐标均归一化至区间[0, 1]。本数据集仅包含一个类别（索引为0）。按照图像级别的80:10:10比例划分为训练集、验证集与测试集，确保单株植株不会同时出现在多个数据划分中。 2. 分类任务数据集：包含35,948株植株级裁剪图像。这些裁剪图像通过将训练好的检测器应用于全部田间图像生成，对每个边界框向外扩展30像素进行填充后，将其调整为固定尺寸的正方形画布，过程中采用保宽高比的信箱填充（letterbox padding）方式。裁剪图像由人工按照叶片取食危害面积占比划分为四个严重度等级：健康（0–10%）、轻度（10–50%）、中度（50–70%）与重度（70–100%）。 数据集目录结构如下： mustard_flea_beetle_dataset/ ├─ detection/ │ ├─ images/ │ │ ├─ train/ （1,282张） │ │ ├─ val/ （160张） │ │ └─ test/ （161张） │ └─ labels/ │ ├─ train/ │ ├─ val/ │ └─ test/ └─ classification/ ├─ healthy/ （32,432张） ├─ mild/ （1,219张） ├─ moderate/ （927张） └─ severe/ （1,370张） 该数据集的类别分布反映了采样田间的自然发生情况，存在严重的类别不平衡问题。四分类严重度划分中存在难以学习的类别边界，将中度与重度等级合并后可显著提升分类模型的训练性能。 数据使用与解读说明： 本次保留了原始的四分类标签，以便用户可以采用自定义的严重度划分方案。若需复现三分类设定，应将中度等级合并至重度等级中。 存在若干影响解读的特性：严重度边界是通过人工目视估算确定，而非通过叶片面积实测得到；标注仅由一名标注人员完成，未评估标注者间一致性；分类裁剪图像带有保宽高比缩放时产生的黑色信箱填充边框；该数据集在其他地区、种植季与芥菜品种上的泛化能力尚未经过验证。 所有图像均为未经过修改的田间实拍照片或由此衍生的裁剪图像，未包含任何合成或数据增强图像。