The Multispectral Potato Plants Images dataset contains aerial images of potato crops and can be utilized for training machine learning models for crop health assessment in precision agriculture applications. It comprises 360 RGB image patches of size 750×750 pixels in JPG format, divided into a train subset of 300 images and a test subset of 60 images. These image patches were extracted from the aerial high-resolution images through cropping, rotating, and resizing operations.

The images were gathered from a field located at the Aberdeen Research and Extension Center - University of Idaho. A Parrot Sequoia multispectral camera mounted on a 3DR Solo drone was employed for image acquisition. Sequoia features several sensors: an RGB sensor with 4,608×3,456 pixels (16 Megapixels) resolution, and four monochrome sensors with 1,280×960 pixels (1.2 Megapixels) resolution capturing narrow bands of light wavelengths: green (550 nm), red (660 nm), red-edge (735 nm), and near-infrared (790 nm). These images were collected as the drone flew over the potato field at a low altitude of 3 meters (≈10 feet). The primary aim was to capture drought stress in Russet Burbank potato plants due to premature plant senescence.

Each image is accompanied by ground-truth annotations (provided as both XML and CSV files) pertaining to regions with healthy and stressed plants, delineated in the images with rectangular bounding boxes. The crop regions were manually labeled into two classes, healthy and stressed, using the graphical annotation software LabelImg. It’s important to note that the testing subset of images is independent of the training subset, meaning that the image patches used for model testing were extracted from different aerial images. The dataset also incorporates corresponding image patches of the spectral sensors with red, green, red-edge, and near-infrared bands, each with a size of 416×416 pixels.

This dataset is well-suited for training machine learning algorithms to assess crop health in a field by detecting healthy vs. stressed plants. The dataset was carefully designed to provide images containing a balanced representation of both healthy and stressed crop classes. However, it’s worth acknowledging that a limitation of the dataset is its relatively small size. Consequently, it is strongly recommended to apply data augmentation and/or transfer learning techniques. For instance, the second dataset mentioned below employs image augmentation techniques, including rescaling pixel intensity, adjusting gamma values for brightness, modifying sigmoid values for contrast, and introducing random noise to the images. By using data augmentation, the initial training set was expanded to 1,500 images, while data augmentation was not applied to the testing set.