Comparative Analysis of Image-Based Deep Learning and Genomic Models for Yield and Protein Content Prediction in Winter Wheat

Xiaoran Chen, Paraskevi Nousi, Mike Boss, Michele Volpi, Lukas Roth; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 7178-7186

We perform a comparative analysis of various image-based and genomic selection deep learning models as a benchmark on the FIP1.0 dataset, a multimodal crop phenotyping dataset offering data collected in multiple years, including time-series imagery, agronomic traits such as yield, protein content, and heading dates, as well as genetic markers and environmental variables including temperature and precipitation. We trained widely used models, including ResNet-50, ConvNeXt, and DINOv2, and evaluated their performance on predictive tasks focused on trait estimation using image sequences and multimodal inputs that integrate imagery with plant height data. Our analysis includes visualizations of temporal feature importance across growth stages and an exploration of learned embeddings to characterize variation across years and genotypes. This benchmark underscores both the challenges and the potential of leveraging heterogeneous data sources for crop modeling, with the goal of advancing genotype-by-environment interaction prediction and interpretation.