Year: 2026
Type: article
Abstract: Reliable crop yield estimation under climate variability requires modeling frameworks that balance predictive accuracy, robustness, and structural consistency across heterogeneous simulation environments. Although deep neural networks (DNNs) have demonstrated strong capability in capturing nonlinear climate–crop relationships, limited attention has been given to systematically comparing architectural strategies that influence stability and uncertainty characterization, particularly under minimum input configurations. This study develops and evaluates a multi-strategy deep learning framework for maize and wheat yield prediction that integrates independent single-task DNNs, multi-task DNNs with shared latent representations, deterministic averaging ensembles, and stochastic ensembles based on Monte Carlo (MC) Dropout. Ethiopia is used as a case study, combining ERA5-derived climate variables with yield simulations from the DSSAT and LPJmL crop models under irrigated and rainfed conditions. Single-task models are trained separately for each crop–regime–simulation combination, while multi-task networks jointly learn structurally related yield targets. Ensemble approaches aggregate model realizations to reduce variance and, in the case of MC Dropout, approximate predictive uncertainty through repeated stochastic inference. Results indicate that ensemble configurations consistently deliver the highest predictive performance, achieving R² values up to 0.94 for wheat and 0.88 for maize, compared to 0.77–0.92 for single-task and 0.76–0.94 for multi-task models. Validation against FAO yield statistics over a 15-year period confirms that ensemble DNNs better reproduce interannual variability and moderate structural divergence between DSSAT- and LPJmL-driven outputs. The findings demonstrate that variance reduction and uncertainty-aware ensemble strategies provide more consistent generalization than architectural task coupling alone, offering a scalable and robustness-oriented framework for climate-sensitive crop yield assessment using parsimonious environmental inputs.
Source: Smart Agricultural Technology
Institutions: DSI Genel Mudurlugu, Wageningen University & Research
Language: English
Subfield: Agronomy and Crop Science
Domain: Life Sciences
SDGs: Zero hunger, Climate action
Open Access status: gold
Estimated APC paid: $1,770
OpenAlex ID: W7204262400