Boundary-Consistent Ordinal Regression for wheat yield prediction using UAV-based multispectral time-series data

Akhtar M.S. Zafar Z. Mahmood Z. Naeem M.K. Yermekbayev K. Fraz M.M.
May 2026 Elsevier B.V.

Computers and Electronics in Agriculture
2026 #246

Accurate plot-scale wheat yield prediction from UAV-based multispectral time-series data is hindered by limited labeled samples, strong phenological variability, and the ordinal nature of yield evaluation in breeding programs. Conventional regression-only objectives minimize numerical error but ignore agronomically meaningful yield categories, often leading to implausible cross-class mispredictions. To address this, we propose Boundary Consistent Ordinal Regression (B-COR), a class-aware multi-task learning framework that jointly predicts continuous yield and ordinal yield classes while explicitly enforcing consistency between predicted yield magnitudes and class boundaries. The B-COR objective integrates three components: a regression loss, ordinal classification supervision, and a boundary-aware modulation term that scales regression penalties based on ordinal deviation. This formulation discourages large cross-category errors while preserving accuracy within yield classes. The framework is evaluated using multi-temporal UAV-derived multispectral data covering key wheat growth stages from booting to maturity, with synthetically generated time-series data employed to mitigate sample scarcity. Robustness and generalizability are assessed through cross-season validation, training on one growing season and testing on an independent season. Compared to a standard mean squared error objective, B-COR reduces RMSE from 1.28 to 0.81 t ha⁻¹ and improves the coefficient of determination (R²) from 0.44 to 0.76, while ordinal classification accuracy increases from 41% to 79%. Notably, severe ordinal mispredictions where plots are assigned to non-adjacent yield categories are reduced from over 30% to approximately 1%. The proposed framework provides a robust, interpretable, and operationally aligned solution for UAV-based yield prediction in breeding trials, supporting more reliable selection and decision-making under real-world agronomic constraints.

Multi-task learning , Ordinal regression , Remote sensing , Yield prediction

Text of the article Перейти на текст статьи

National University of Sciences and Technology (NUST), Islamabad, Pakistan
National Agricultural Research Center, Islamabad, Pakistan
Institute of Biotechnology and Ecology, Zhetysu University, Taldykorgan, Kazakhstan
University of Staffordshire, College Rd, Stoke-on-Trent, ST4 2DE, United Kingdom

National University of Sciences and Technology (NUST)
National Agricultural Research Center
Institute of Biotechnology and Ecology
University of Staffordshire

10 лет помогаем публиковать статьи Международный издатель

Книга Публикация научной статьи Волощук 2026 Book Publication of a scientific article 2026