Grant: Discovery Grants Program - Individual

Recent progress in artificial intelligence, embedded systems, and robotics created an incredible potential to transform agriculture. Worldwide we see collaborative efforts between farmers, plant breeders, and plant researchers on one side and machine learning (ML) experts on the other side. The goal is often to detect or classify objects in image data with the help of deep neural networks. These models are ideally trained on a sample set that covers all the variance to be encountered under field conditions. The brute-force approach to obtain this variety is to create a dataset of immense size, such that even the unlikely cases are eventually captured. However, this data-hungry approach creates overhead in the collection, sharing, and storing of data, and in the training time of models. Generating such large datasets is particularly difficult in agricultural environments. First, the data that must be collected can vary significantly (just two examples are detecting weeds in a soybean field versus counting diseased kernels of wheat plants); second the data is collected in locations with limited IT infrastructure and internet connectivity and is often coupled to specific growth stages resulting in short time windows for collection; third, the personnel often does not have a ML background and limited experience in dataset requirements; fourth, data collection in the field is time- and cost-intensive, due to using expensive imagers or the involvement of farming vehicles. The need of large amounts of data for ML models is thus in conflict with the difficulties of obtaining field datasets. We must ask how data quantity and quality influences model performance, particularly in agricultural applications, where data-related costs are substantially increased. Methods to create small, yet rich, datasets must be found; further these methods must be robust and transferable since application scenarios vary significantly. In addition, feedback between data collection and model training must propagate quickly to allow for corrections in short collection periods and to avoid collecting unsuitable, redundant, or stale data. In this research program we tackle the challenges of data collection from several angles. First, we develop new insights and heuristics for predicting the number of samples required to train a model to a target performance. Second, we investigate how datasets can be pruned and refined into smaller training sets with near equal data value. Third, we develop a software framework that facilitates short feedback loops between model training and data collection processes - thus enhancing data collection online. Fourth, we develop found solutions for edge devices close to the data sources. Special considerations must be taken as these edge devices can only support lightweight implementations. Results of this research will enable practitioners, such as farmers, breeders, and plant researchers to collect better data easier.