Date of Publication
4-2026
Document Type
Master's Thesis
Degree Name
Master of Science in Applied Physics
Subject Categories
Physics
College
College of Science
Department/Unit
Physics
Thesis Advisor
Edgar A. Vallar
Defense Panel Chair
Maria Cecilia D. Galvez
Defense Panel Member
Jazzie R. Jao
Jumar G. Cadondon
Abstract (English)
Dragon fruit stem canker represents a major threat to crop productivity in the Philippines. Traditional diagnostic methods rely on manual visual inspection, which is often subjective and inefficient for early-stage detection. The study established a multi-scale diagnostic framework that integrates ground-level smartphone imagery with canopy-level aerial multispectral data to improve disease monitoring.
The ground-based diagnostic component utilized a specialized three-stage deep learning pipeline. YOLOv8m for high-precision cladode localization (91.9% mAP), Segment Anything Model 2 (SAM2) for instance segmentation, and a hybrid VGG16-Vision Transformer (ViT) for severity classification.
Simultaneously, multispectral drone flights were conducted at 30-meter and 50-meter altitudes to evaluate canopy health through various vegetation indices. Validation against high-precision spectrometer data confirmed that consumer-grade smartphone sensors effectively track physical reflectance trends (r=0.83) and identified the modified Photochemical Reflectance Index (mPRI) as a robust cross-scale indicator.
Results showed that the hybrid VGG16-ViT model achieved perfect F1-scores (100%) for the Healthy and Severe classes, demonstrating outstanding standalone convolutional architecture performance in identifying subtle early-stage lesions. Aerial analysis indicated that 50-meter flight altitude offers a “spectral smoothing effect” optimal for broad screening (r=0.986), while 30-meter flight altitude provides superior categorical agreement with ground truth (κ=0.661). This multi-scale approach enables farmers to bridge detailed individual-stem analysis with scalable canopy monitoring to deliver targeted, sustainable interventions.
Abstract Format
html
Abstract (Filipino)
Ang stem canker sa dragon fruit ay isang malaking banta sa produksyon ng pananim sa Pilipinas. Ang mga nakasanayang pamamaraan ng pagsusuri ay nakadepende sa manwal na pag-inspeksyon gamit ang paningin, na kadalasang subhektibo at hindi epektibo para sa pagtukoy ng sakit sa maagang yugto nito. Ang pag-aaral na ito ay bumuo ng isang multi-scale diagnostic framework na pinagsasama ang mga imahe mula sa smartphone sa ground-level at ang aerial multispectral data sa antas ng canopy upang mapabuti ang pagmomonitor ng sakit.
Ang bahaging ground-based diagnostic ay gumamit ng espesyalisadong three-stage deep learning pipeline. Ginamit ang YOLOv8m para sa mataas na katumpakan ng cladode localization (91.9% mAP), Segment Anything Model 2 (SAM2) para sa instance segmentation, at isang hybrid na VGG16-Vision Transformer (ViT) para sa pag-uuri ng kalubhaan ng sakit (severity classification).
Kasabay nito, nagsagawa ng mga multispectral drone flight sa taas na 30-metro at 50-metro upang suriin ang kalusugan ng canopy sa pamamagitan ng iba't ibang vegetation indices. Pinatunayan ng balidasyon laban sa high-precision spectrometer data na ang mga consumer-grade smartphone sensor ay epektibong nakakasubaybay sa mga physical reflectance trend (r=0.83) at natukoy ang modified Photochemical Reflectance Index (mPRI) bilang isang matibay na cross-scale indicator.
Ipinakita ng mga resulta na ang hybrid VGG16-ViT model ay nakakuha ng perpektong F1-scores (100%) para sa mga kategoryang Healthy (Malusog) at Severe (Malala), na nagpapakita ng natatanging pagganap ng standalone convolutional architecture sa pagtukoy ng mga banayad at nasa maagang yugtong sugat ng halaman (subtle early-stage lesions). Ipinahiwatig ng pagsusuring aerial na ang 50-metrong taas ng paglipad ay nag-aalok ng isang “spectral smoothing effect” na pinakamainam para sa malawakang pagsusuri o broad screening (r=0.986), habang ang 30-metrong taas ay nagbibigay ng mas mahusay na categorical agreement sa ground truth (κ=0.661). Ang multi-scale na pamamaraang ito ay nagbibigay-daan sa mga magsasaka na pag-ugnayin ang detalyadong pagsusuri sa bawat tangkay (individual-stem analysis) at ang malawakang pagmomonitor ng canopy upang makapaghatid ng mga tiyak at napapanatiling interbensyon.
Abstract Format
html
Language
English
Format
Electronic
Keywords
Fruit—Diseases and pests; Deep learning (Machine learning); Remote sensing
Recommended Citation
Joromo, A. N. (2026). Early detection of dragon fruit stem canker using a VGG16-vision transformer hybrid model for ground imagery and aerial multispectral analysis. Retrieved from https://animorepository.dlsu.edu.ph/etdm_physics/29
Upload Full Text
wf_yes
Embargo Period
4-17-2029
