<p>Accurate crop yield prediction is vital for food security and sustainable agricultural planning, particularly under increasing climate variability. However, most existing approaches lack adaptive preprocessing and rely on manually tuned models with limited generalization. This study aims to address these gaps by integrating large language model (LLM)-informed preprocessing with neural modeling and intelligent optimization. Using Mixtral-8x7B-Instruct, domain-specific features were engineered and refined via principal component analysis and KMeans clustering. Among baseline models, neural ordinary differential equation (NODE) performed best (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$ \text {RMSE} {= 0.1109} $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>RMSE</mtext> <mrow> <mo>=</mo> <mn>0.1109</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$ {R}^{{2}} {= 0.8894} $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mrow> <mi>R</mi> </mrow> <mn>2</mn> </msup> <mrow> <mo>=</mo> <mn>0.8894</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>). Metaheuristic tuning using ten algorithms further improved accuracy, with the Football Optimization Algorithm (FbOA) achieving the highest gains (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$ \text {RMSE} {= 0.0232} $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>RMSE</mtext> <mrow> <mo>=</mo> <mn>0.0232</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$ {R}^{{2}} {= 0.9850} $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mrow> <mi>R</mi> </mrow> <mn>2</mn> </msup> <mrow> <mo>=</mo> <mn>0.9850</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$ \text {NSE} {= 0.9740} $</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mtext>NSE</mtext> <mrow> <mo>=</mo> <mn>0.9740</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>). The proposed framework demonstrates how combining LLM-guided feature design and optimization-enhanced neural models offers a scalable, high-precision solution for agricultural forecasting.</p>

Agricultural Yield Forecasting Using LLM-Informed Preprocessing and Neural ODE Models Optimized via the Football Optimization Algorithm

期刊论文
发表日期：

2026年1月16日

查看全文

Mai Alduailij,
Sayed Elkenawy,
Amel Ali Alhussan,
Ebrahim A. Mattar,
Amal H. Alharbi,
Doaa Sami Khafaga,
Safaa Zaman,
Marwa M. Eid

摘要

Accurate crop yield prediction is vital for food security and sustainable agricultural planning, particularly under increasing climate variability. However, most existing approaches lack adaptive preprocessing and rely on manually tuned models with limited generalization. This study aims to address these gaps by integrating large language model (LLM)-informed preprocessing with neural modeling and intelligent optimization. Using Mixtral-8x7B-Instruct, domain-specific features were engineered and refined via principal component analysis and KMeans clustering. Among baseline models, neural ordinary differential equation (NODE) performed best ( $ \text {RMSE} {= 0.1109} $ $RMSE = 0.1109$ , $ {R}^{{2}} {= 0.8894} $ $R^{2} = 0.8894$ ). Metaheuristic tuning using ten algorithms further improved accuracy, with the Football Optimization Algorithm (FbOA) achieving the highest gains ( $ \text {RMSE} {= 0.0232} $ $RMSE = 0.0232$ , $ {R}^{{2}} {= 0.9850} $ $R^{2} = 0.9850$ , $ \text {NSE} {= 0.9740} $ $NSE = 0.9740$ ). The proposed framework demonstrates how combining LLM-guided feature design and optimization-enhanced neural models offers a scalable, high-precision solution for agricultural forecasting.