A comprehensive dataset for monitoring germination of Cannabis sativa in greenhouse-controlled environments

Abstract

Germination monitoring in crucial: it aids in planning and maximizing crop yields, contributing to sustainable agriculture practices. Furthermore, in the cultivation of Cannabis sativa L., precise control over the number of plants during the germination process is critical due to legal and regulatory restrictions. To assist growers in tracking germination progress, we planned to developed a classifier that informs them precisely when a plant has successfully germinated in the seedbed. A well-curated dataset is essential for teaching the algorithm to recognize different features. The dataset should encompass a diverse range of examples to ensure that the classifier can accurately identify and categorize instances it encounters during its operation. The quality and diversity of the dataset play a pivotal role in the performance and reliability of the developed classifier. Due to the limited information available on cannabis crops, we undertook the task of constructing an image dataset from the ground up. This dataset was meticulously crafted through a rigorous process. To build the dataset, we sow two varieties of cannabis, Finola and Kompolti, in separate seed trays, each containing 72 cells, in a greenhouse. A camera fixed above the seedbeds took one image every hour throughout the germination experiment. Four iterations were carried out: two without controlling climatic conditions, and two with controlled conditions and photoperiod. Then, we cropped the images and applied a homography process to correct perspective. The resulting images of each cell for each date and hour were labelled using six categories. first, the images were labelled considering the categories germination, nongermination, only cotyledon, and true leaves, then, they were labelled again using the categories invasion and non-invasion. The result was a comprehensive dataset comprising 80,640 images of seedbed cells showcasing plants at various growth stages. This paper outlines the step-by-step process employed in creating this image dataset.