Field-level modeling independent of spatial extent

Develop crop yield prediction models capable of operating at the field level independent of field spatial extent, rather than relying solely on pixel-level approaches constrained by variable field sizes.

Background

Most existing approaches in the paper’s benchmarks treat each pixel as an independent sample because fields vary widely in size and shape, and data availability can be limited. While some models capture local neighborhood information, a general solution that handles entire fields regardless of their spatial extent has not been established.

The authors identify this gap explicitly as an open challenge, highlighting the need for models that can coherently represent and predict yield for whole fields despite variability in their geometry.