Towards precision agriculture: metaheuristic model compression for enhanced pest recognition

• Published in: Scientific reports Vol. 15; no. 1; pp. 20805 - 13
• Main Authors: Parez, Sana, Alghamdi, Norah Saleh, Mahmood, Tahir, Ullah, Waseem, Khan, Muhammad Attique, Houda, Taha, Dilshad, Naqqash
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 01.07.2025; Nature Portfolio
• Subjects: 639/705/117; 639/705/258; Agriculture - methods; Algorithms; Animals; Convolution neural network; Crops, Agricultural - parasitology; Deep Learning; Disease recognition; Humanities and Social Sciences; Machine learning; multidisciplinary; Pest detection; Plant Diseases; Precision agriculture; Science; Science (multidisciplinary); Deep learning; Precision agriculture; Convolution neural network; Disease recognition; Machine learning; Pest detection
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-08307-5

Crop diseases and insect pests pose significant challenges to agricultural productivity, often resulting in considerable yield losses. Traditional pest recognition methods, which rely heavily on manual feature extraction, are not only time consuming and labor intensive but also lack robustness in diverse conditions. While deep learning (DL) models have improved performance over conventional approaches, they typically suffer from high computational demands and large model sizes, limiting their real-world applicability. This study proposes a novel and efficient DL-based framework for the accurate identification and classification of crop pests and diseases. The core of this approach integrates InceptionV3 as a backbone feature extractor to capture rich and discriminative features, enhanced further using a channel attention (CA) mechanism for feature refinement. To reduce model complexity and improve deployment feasibility, a metaheuristic optimization algorithm was incorporated that significantly reduces computational overhead without compromising performance. The proposed model was rigorously evaluated on the CropDP-181 dataset, outperforming several state-of-the-art methods in both classification accuracy and computational efficiency. Notably, the proposed method achieved a precision of 0.932, recall of 0.891, F1-score of 0.911, an overall accuracy of 88.50%, and an MCC of 0.816 demonstrating its effectiveness and practical potential in real-time agricultural monitoring systems.