| Titre : | Hybrid artificial intelligence for climate-smart wheat production: optimizing resource efficiency and carbon reduction from Mediterranean to Nordic regions (2026) |
| Auteurs : | M.F. Rabbi |
| Type de document : | Article |
| Dans : | Cleaner Environmental Systems (vol. 21, June 2026) |
| Article en page(s) : | p. 100446 |
| Langues : | Anglais |
| Langues du résumé : | Anglais |
| Catégories : |
Thésaurus IAMM CEREALICULTURE ; BLE ; PRODUCTION AGRICOLE ; CARBONE ; GESTION DES RESSOURCES ; EPUISEMENT DES RESSOURCES ; INTELLIGENCE ARTIFICIELLE ; AIDE A LA DECISION ; METHODE D'OPTIMISATION ; REGION MEDITERRANEENNE |
| Résumé : | Climate change intensifies agricultural resource demands across European wheat production regions through weather extremes and water scarcity, necessitating decision support systems that optimize multiple sustainability objectives under diverse climate scenarios. This study developed and validated a hybrid artificial intelligence framework integrating expert systems (15 production rules), fuzzy logic controllers, four reinforcement learning algorithms (Q-learning, Deep Q-Network, Policy Gradient, Actor-Critic), and 118 autonomous agents for climatesmart wheat management. Performance evaluation employed 30 independent growing seasons across Mediterranean, Central European, and Nordic regions under normal, drought, and extreme heat scenarios, with empirical validation from 2015 to 2024. Under normal conditions, the system achieved 7.8% yield improvements, 17.8% water savings, and 15.1% energy reductions relative to conventional management. Benefits amplified under climate stress; drought scenarios revealed 34.8% yield gains, 29.1% water savings, and 18.3% energy reductions, while extreme heat conditions yielded 26.1% yield improvements, 29.4% water savings, and 22.1% energy efficiency gains. Actor-Critic reinforcement learning achieved cumulative rewards of 519.7 versus 409.6 for Qlearning, demonstrating superior temporal optimization. Sequential ablation confirmed that withheld reinforcement learning produced the largest single-component performance decline (hybrid F1-score from 0.893 to 0.831), while removal of any component reduced hybrid performance, confirming that gains arise from complementary multi-paradigm contributions. Regional analysis confirmed spatial transferability across diverse pedoclimatic contexts. Empirical validation yielded prediction accuracies of R2 = 0.802 for yield and R2 = 0.783 for water use, with carbon footprint reductions of 18.7-20.3% across climate scenarios; these simulation-derived findings motivate prospective field trials in resource-efficient wheat production systems under increasing climate variability. |
| Cote : | En ligne |
| URL / DOI : | https://doi.org/10.1016/j.cesys.2026.100446 |
Accueil
Sélection de la langue
Adresse
ÉQUIPE
Nous contacter
Informations pratiques
contact
