A Dataset for Early Detection of Corn Leaf Pests in Precision Agriculture

Corn plantations were visited to collect images of both infected and healthy leaves. An agricultural expert labeled the samples as follows: "S" for images containing only healthy leaves, "M1" for leaves infected by Spodoptera frugiperda, "M2" for leaves infected by Helminthosporium leaf blight, and "M3" for leaves infected by Zonocerus variegatus.

The original dimensions of the images varied: 4248 × 5664, 1920 × 2560, 2448 × 3264, 3120 × 4160, and 2448 × 3264. All the images were resized to 400 × 400 pixels using Python programming. The resizing process maintained the quality of the images. In total, 1308 images were collected from various fields, resulting in a dataset containing healthy leaves, and leaves infected by Spodoptera frugiperda, Helminthosporium, and Zonocerus variegatus. Each image was labeled according to the following format: field_code – image_condition – growth_stage – date_taken – sequence_number. This coding was implemented to facilitate the tracking of the state of each corn leave within the dataset.

Note that we published a paper titled "Automatic Segmentation Based on the CNDVI (Combined Normalized Difference Vegetation Index)" (see [1]) using an initial, small version of this dataset. The work in [1] focuses on background segmentation of corn plants in the field. The algorithm proposed in [1] was further applied to the dataset to segment the background of corn by removing all background elements, leaving only the corn leaf in the image. Another version of background segmentation is also proposed in this paper, using manual segmentation. Additionally, we introduce a dataset that highlights all infected areas on the corn leaf.

Overall, we propose 8 datasets described as follows:
Dataset 1: Natural images
Dataset 2: Images with manually segmented backgrounds
Dataset 3: Images with automatically segmented backgrounds by CNDVI
Dataset 4: Augmented version of Dataset 1 by a factor of 9
Dataset 5: Augmented version of Dataset 2 by a factor of 9
Dataset 6: Augmented version of Dataset 3 by a factor of 9
Dataset 7: Images of infected leaves with manually segmented backgrounds and infected areas highlighted
Dataset 8: Augmented version of Dataset 7 by a factor of 9

本研究通过走访玉米种植园，采集染病与健康玉米叶片的图像样本。由农业专家对样本进行标注，规则如下：仅包含健康叶片的图像标注为"S"；感染草地贪夜蛾（Spodoptera frugiperda）的叶片图像标注为"M1"；感染长蠕孢叶枯病（Helminthosporium leaf blight）的叶片图像标注为"M2"；感染杂色斑翅蝗（Zonocerus variegatus）的叶片图像标注为"M3"。

图像原始分辨率尺寸存在差异，涵盖4248×5664、1920×2560、2448×3264、3120×4160及2448×3264多种规格。本研究通过Python编程将所有图像统一调整至400×400像素，且调整过程未对图像画质造成损耗。本研究从多个试验田中共采集到1308张图像，最终构建的数据集涵盖健康叶片、感染草地贪夜蛾、长蠕孢叶枯病及杂色斑翅蝗的玉米叶片图像。每张图像均按照以下格式进行命名标注：地块编号-图像状态-生育时期-拍摄日期-序列编号。该编码规则旨在便于追踪数据集中每一片玉米叶片的相关状态信息。

需说明的是，本团队曾基于该数据集的初始小样本版本，发表了题为"Automatic Segmentation Based on the CNDVI (Combined Normalized Difference Vegetation Index)"的论文（详见文献[1]）。文献[1]的研究聚焦于田间玉米植株的背景分割任务，文中提出的算法被进一步应用于本数据集，通过移除所有背景元素、仅保留玉米叶片区域，完成玉米图像的背景分割。本文同时提出了另一种背景分割方案，即人工分割法。此外，本研究还构建了一个标注玉米叶片所有染病区域的数据集。

综上，本研究共构建8个数据集，详情如下：
数据集1：原始自然图像
数据集2：经人工分割背景的图像
数据集3：基于复合归一化植被指数（CNDVI，Combined Normalized Difference Vegetation Index）实现自动背景分割的图像
数据集4：数据集1经9倍数据增强后的版本
数据集5：数据集2经9倍数据增强后的版本
数据集6：数据集3经9倍数据增强后的版本
数据集7：经人工分割背景且标注染病区域的染病叶片图像
数据集8：数据集7经9倍数据增强后的版本