Publikation
Informed Learning for efficient Crop Yield Prediction
Miro Miranda Lorenz; Francisco Mena; Marcela Charfuelan Oliva; Andreas Dengel (Hrsg.)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS-2025), August 3-8, Brisbane, Australia, IEEE, 2025.
Zusammenfassung
Crop yield prediction at field and subfield level is a major challenge in Earth Observation (EO), supporting decision-making and global food security. The rapid increase in EO data availability has driven the development of machine learning models for crop yield prediction, capable of processing EO time series data without requiring extensive preprocessing and expert knowledge. However, models such as Transformers are computationally expensive because of their quadratic cost function of the attention mechanism, consequently increasing computational burdens and the overall carbon footprint.
Informed ML (iML) offers a promising solution by integrating prior knowledge into the learning algorithm, thereby reducing computational demands without compromising accuracy.
In this research, we investigate the effectiveness of iML for efficient crop yield prediction. We compare various time series preprocessing schemes on four yield datasets and present a method that samples a single time step for each crop growth stage. Thereby, reducing the number of required time steps and training time without sacrificing performance. Additionally, we demonstrate interesting insights into meaningful agronomic patterns.