An explainable SR-GAN and MobileNetV2 framework for aerial pest detection in precision agriculture
Computers and Electronics in Agriculture · 2026-09 · 252 · 112097
DOI: 10.1016/j.compag.2026.112097 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computers and Electronics in Agriculture, 2026-09. Verfasst von Sundarrajan Munusamy und 6 weiteren Autor:innen. Die Arbeit nennt 51 Literaturangaben. Gefördert von: Princess Nourah bint Abdulrahman University.
Abstract (Originalsprache)
As pest problems in agriculture continue to grow, there is now greater demand for systems that can quickly spot them. It is common to use human scouts for spotting pests or standard machine-learning techniques, but they cannot easily process the low-resolution images taken by Unmanned Aerial Vehicles(UAVs). Using blurred or poor quality data in deep learning usually causes the model to miss some items and select the incorrect choice. To deal with the mentioned problems, this research combines Super-Resolution Generative Adversarial Network (SR-GAN) technology with a lightweight classifier called MobileNetV2 for more effective monitoring of aerial pests. The main function of the SR-GAN module is to transform UAV images with low resolution into ones with high resolution that keep important details of each pest. After improving the images, MobileNetV2 is used to classify them effectively and accurately. Trust in the framework is further increased by Grad-CAM which makes it clear which features of the input affect the prediction of the model. Images created using super-resolution are deemed visually better due to the boost in PSNR, SSIM and MSE. Classification tests show that the improved models have superior accuracy, graphs and analysis of important features. The method improved upon the base CNN by 18%, reaching a 92.6% rate of correct predictions. Moreover, the test results indicated that PSNR values were 5.2 dB higher than with the CNN. The proposed method has more consistent outputs and better behavior under varied farming conditions with Accuracy = 92.6%, Precision = 93.1%, Recall = 92.2%, F1 Score = 92.4%, Parameters = 2.3 M, FLOPs = 0.3G, Inference Time = 30 ms, which can effectively detect the pests even in occluded, shadow, or low-visibility cases. It combines various approaches and remedies the existing system problems, offering flexibility for immediate use in agriculture to help save crops and make pesticide use more efficient.
Förderung
- Princess Nourah bint Abdulrahman University
Zitieren
APA
Munusamy, S., Jothi, A., Kosalram, T., Dhanaraj, R. K., Mon, F. A., Sivakumar, N. R., & Alkhayyat, A. (2026). An explainable SR-GAN and MobileNetV2 framework for aerial pest detection in precision agriculture. Computers and Electronics in Agriculture, 252, 112097. https://doi.org/10.1016/j.compag.2026.112097
Vancouver
Munusamy S, Jothi A, Kosalram T, Dhanaraj RK, Mon FA, Sivakumar NR, et al. An explainable SR-GAN and MobileNetV2 framework for aerial pest detection in precision agriculture. Computers and Electronics in Agriculture. 2026;252:112097. doi:10.1016/j.compag.2026.112097
BibTeX
@article{Munusamy2026,
author = {Munusamy, Sundarrajan and Jothi, Akshya and Kosalram, Thamaraiselvi and Dhanaraj, Rajesh Kumar and Mon, Feslin Anish and Sivakumar, Nithya Rekha and Alkhayyat, Ahmad},
title = {An explainable SR-GAN and MobileNetV2 framework for aerial pest detection in precision agriculture},
journal = {Computers and Electronics in Agriculture},
year = {2026},
volume = {252},
pages = {112097},
doi = {10.1016/j.compag.2026.112097}
}
EUVIMED