Date of Publication

11-2024

Document Type

Master's Thesis

Degree Name

Bachelor of Science in Computer Engineering - Ladderized

Subject Categories

Electrical and Computer Engineering

College

Gokongwei College of Engineering

Department/Unit

Electronics And Communications Engg

Thesis Advisor

Ann E. Dulay

Defense Panel Chair

Melvin K. Cabatuan

Defense Panel Member

Cesar A. Llorente

Dino Dominic F. Ligutan

Abstract (English)

Viticulture, or grape cultivation, is an emerging industry in the Philippines, benefitting from an extended growing season due to the lack of winter. However, the rainy season also brings pests, like fall armyworms and Chinese rose beetles, that devastate crops. Overusing pesticides to control these insects incurs high costs to farmers, increases pest resistance, and poses environmental risks. Pest recognition models provide an avenue for controlling the application of pesticides. Unfortunately, current pest detection studies overlook environmental factors that affect image quality, like rain.

This study introduces rain augmentation for pest detection by adapting the unsupervised deraining model DerainCycleGAN. A clear and rainy dataset featuring the most common insects in a vineyard in San Juan, Batangas, was collected to initially evaluate the model's effectiveness. When scaled to the benchmark dataset IP102, rain augmentation improved performance on clear images as well. Compared with YOLOv8m, used as a baseline by recent state-of-the-art models, rain augmentation improved the mAP50 by 1.3%, achieving the second-best 63.0% mAP50 without significant architectural modifications. Rain augmentation is also robust to downscaling, achieving a 1.4% mAP50 improvement over the 320x320 baseline. Additional testing with recently introduced models proved promising, further boosting performance to 64.8% mAP50 using YOLOv11m trained on the augmented dataset.

DerainCycleGAN-based augmentation's performance on a separate corrupted test set was observed by using a modified dataset split of IP102 and updating the baseline model to YOLOv8s for the reduced training data. Across various severities of weather degradation, the augmented model exceeds baseline YOLOv8s by up to 8.6% mAP50, even outperforming unaugmented YOLOv8m trained on the original split in terms of robustness. Adopting such techniques offers a promising way to address data scarcity and limited dataset diversity in pest detection, enabling the development of more effective models for real-world environments and promoting sustainable pest control.

Abstract Format

html

Abstract (Filipino)

None

Abstract Format

html

Language

English

Format

Electronic

Keywords

Computer vision

Upload Full Text

wf_yes

Embargo Period

11-2025

Download

Share

COinS