Potato Disease Detection via Joint Feature Selection and

Hyperparameter Optimization of the Feature Tokenizer

Transformer with iHOW

Doaa Sami Khafaga1,*

1 Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University,

P.O. Box 84428, Riyadh 11671, Saudi Arabia

Received: January 12, 2026 Revised: March 15, 2026 Accepted: May 19, 2026 ⋆ Corresponding author

ABSTRACT

Potato leaf diseases represent a persistent threat to global food security, demanding diagnostic systems that are

not only accurate but also computationally efficient and deployable in resource-constrained agricultural

environments. This study introduces a unified framework that integrates the Feature Tokenizer Transformer

(FT-Transformer) with the Improved iHOW Optimization Algorithm, a metaheuristic designed for simultaneous

feature selection and hyperparameter tuning. The FT-Transformer alone provides a competitive baseline,

achieving 81.61% accuracy and an F1-score of 81.45%. Incorporating iHOW-based feature selection raises

accuracy to 90.90% while reducing redundancy and enhancing precision, recall, and generalization. A final stage

of iHOW-guided hyperparameter optimization further improves performance to 98.35% accuracy and 98.33%

F1-score, outperforming all comparative metaheuristic variants. Beyond predictive gains, the optimized model

demonstrates strong efficiency, requiring only 12.45 s of training time and 256.8 MB of memory. Taken together,

these results establish iHOW+FT-Transformer as a scalable, interpretable, and resource-conscious solution for

real-time plant disease classification. The proposed pipeline advances the state of precision agriculture by

enabling robust, low-overhead crop monitoring systems adaptable to diverse and low-resource farming contexts.

Keywords: Potato Disease Detection FT-Transformer iHOW Optimization Feature Selection Hyperparameter

Tuning Precision Agriculture

1. INTRODUCTION

The global demand for sustainable and intelligent agriculture

has intensified in recent years, driven by climate change, population

growth, and the urgent need to safeguard food security.

Among the most persistent threats to agricultural productivity

are crop diseases, which reduce both yield quantity and

quality and cause significant economic and social disruption—

particularly in regions with limited access to timely

diagnostic resources [1, 2]. Potatoes (Solanum tuberosum),

a major staple crop worldwide, are especially vulnerable to

diverse bacterial and fungal infections that can trigger severe

losses if not detected and managed promptly [3].

Effective disease control depends on early-stage diagnosis

and targeted intervention. Conventional diagnostic practices,

often based on manual inspection by trained experts, are constrained

by scalability, speed, and consistency [4]. These

limitations are magnified in large-scale farms and resourceconstrained

rural settings where expert access is scarce. To