Global Crop Yields Forecasting Using Informer Architecture

Optimized with Football-Inspired Metaheuristic

Toufik Mzili1,*

1 Faculty of Sciences, Chouaib Doukkali University, El Jadida, Morocco

Received: January 03, 2026 Revised: March 01, 2026 Accepted: May 04, 2026 ⋆ Corresponding author

ABSTRACT

Accurate forecasting of agricultural production is essential for managing food supply chains, guiding policy decisions,

and building resilience against climate variability. However, modeling long-range, country-level crop production

remains challenging due to temporal complexity, nonlinear dependencies, and the need for highly generalizable

prediction systems. This study addresses these challenges by developing a hybrid forecasting framework that

combines deep learning architectures with metaheuristic hyperparameter optimization. A global dataset spanning

tomato and potato production from 1961 to 2021 was used to evaluate multiple forecasting models, including

Informer, N-BEATS, LogTrans, N-HITS, EALSTM, TST, and LSTM. The Informer model achieved the best baseline

performance (RMSE = 0.0799; NSE = 0.9070) and was selected for optimization. A comparative analysis was

conducted using several metaheuristic algorithms, with particular focus on the Football Optimization Algorithm

(FbOA), a novel strategy inspired by cooperative team dynamics. FbOA delivered the highest gains across all

evaluated metrics, reducing RMSE to 0.00109, MSE to 1.19×10−6, and increasing NSE to 0.9260, R2 to 0.9600,

and the correlation coefficient r to 0.9550. These results confirmed that metaheuristic tuning substantially enhances

the forecasting capability of deep models, particularly when guided by domain-inspired search logic. The proposed

framework demonstrates strong potential for integration into real-time, AI-driven agricultural decision support

systems, offering scalable solutions for food security planning, climate-smart agriculture, and long-term sustainability

forecasting.

Keywords: Crop Yield Prediction Neural Ordinary Differential Equation (NODE) Metaheuristic Optimization

Football Optimization Algorithm (FbOA) Large Language Models (LLMs)

1. INTRODUCTION

The predictive agricultural production is at the center of the

world sustainability, economic stability, and food security

initiatives. In a world where climatic volatility and population

increase are becoming the new norm and resource

competition is a reality, then skills in accurate prediction of

crop yields are no longer optional or recommended but increasingly

mandatory. Among the myriad of crops cultivated

worldwide, Solanum tuberosum (potato) and Solanum lycopersicum

(tomato) are of exceptional strategic importance[1, 2].

Potatoes form a staple of caloric diet in both normal and developing

economies, and tomatoes are both a market product

and a processed product traded on the world market. Their

economic, nutritional, and agronomic interdependence makes

them a perfect choice for an item of a high-resolution, datadriven

forecasting system[3, 4].

With the collection by decades of structural agricultural data

by organizations such as the Food and Agriculture Organization

(FAO), there has been a wealth of new opportunities