A Decision-Consistent Framework for Yield Prediction and Genotype Selection using UAV-Based Temporal Phenotyping

Modern plant breeding depends on identifying superior genotypes quickly and reliably. UAV-based high-throughput phenotyping and machine learning have improved yield estimation, but most approaches treat it as conventional numerical regression, which overlooks what breeding programs actually need: reliable ranking of candidate lines, consistency at advancement thresholds and robustness when field data is scarce. The thesis proposes a unified decision-aware spatio-temporal learning framework combining a registration and harmonization pipeline that keeps multi-date UAV orthophotos geometrically consistent, a yield-conditioned synthetic time-series generator that offsets limited field data, a Boundary-Consistent Ordinal Regression (B-COR) formulation that embeds agronomic performance tiers into the learning objective, and a Rank-Aware Multi-Task Learning (RA-MTL) model that jointly optimizes regression, ordinal classification and ranking. Evaluated on sunflower, wheat and rice across independent seasons, it achieves cross-crop R² of 0.62–0.77, Spearman rank correlations of 0.85–0.88 and Top-20% selection accuracy of 0.77–0.80, advancing UAV-based phenotyping from isolated yield prediction toward selection-centric intelligence.