A novel automated cloud-based image datasets for high throughput phenotyping in weed classification

Deep learning-based weed detection data management involves data acquisition, data labeling, model development, and model evaluation phases. Out of these data management phases, data acquisition and data labeling are labor-intensive and time-consuming steps for building robust models. In addition, low temporal variation of crop and weed in the datasets is one of the limiting factors for effective weed detection model development. This article describes the cloud-based automatic data acquisition system (CADAS) to capture the weed and crop images in fixed time intervals to take plant growth stages into account for weed identification. The CADAS was developed by integrating fifteen digital cameras in the visible spectrum with gphoto2 libraries, external storage, cloud storage, and a computer with Linux operating system. Dataset from CADAS system contain six weed species and eight crop species for weed and crop detection. A dataset of 2000 images per weed and crop species was publicly released. Raw RGB images underwent a cropping process guided by bounding box annotations to generate individual JPG images for crop and weed instances. In addition to cropped image 200 raw images with label files were released publicly. This dataset hold potential for investigating challenges in deep learning-based weed and crop detection in agricultural settings. Additionally, this data could be used by researcher along with field data to boost the model performance by reducing data imbalance problem.