Implementasi Dan Analisis Kinerja Model Hibrida Transfer Learning VGG19 Dan K-Nearest Neighbor Untuk Klasifikasi Penyakit Tanaman Bunga Matahari

Azzahra, Ratna Amalia (2026) Implementasi Dan Analisis Kinerja Model Hibrida Transfer Learning VGG19 Dan K-Nearest Neighbor Untuk Klasifikasi Penyakit Tanaman Bunga Matahari. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026221209-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Tanaman bunga matahari (Helianthus annuus L.) merupakan komoditas pertanian bernili ekonomi tinggi yang rentan terhadap berbagai penyakit, seperti Downy mildew dan Gray mold yang dapat menyebabkan penurunan hasil panen secara signifikan apabila tidak terdeteksi sejak dini. Keterbatsan kemampuan identifikasi penyakit secara visual oleh petani serta minimnya akses terhadap tenaga ahli agronomi di lapangan mendorong kebutuhan akan sistem deteksi penyakit tanaman yang cerdas, cepat, dan dapat dioperasikan melalui perangkat seluler. Tugas akhir ini bertujuan untuk membangun sistem klasifikasi penyakit tanaman bunga matahari berbasis arsitektur hibrida yang menggabungkan Transfer learning VGG19 sebagai pengekstraktor fitur visual dengan algoritma KNN sebagai pengklasifkasi akhir. Dataset yang digunakan terdiri dari 2.392 citra digital tanaman bunga matahari yang mencakup lima kategori kondisi, yaitu Downy mildew, Fresh Flower, Fresh leaf, Gray mold, dan Leaf scars yang dibagi secara terstratifikasi dengan rasio 70:15:15 menjadi 1.674 training set, 359 validation set, dan 359 citra testing set. Model VGG19 dikonfigurasi menggunakan strategi fine-tuning parsial dengan melonggarkan 14 lapisan terakhir menggunakan laju pembelajaran senilai 5e-6 sehingga dapat menghasilkan akurasi validasi optimal sevesar 91.26% pada epoch ke-15. Vektor fitur berdimensi 512 yang diekstrak dari lapisan GlobalAveragePooling2D kemudian digunakan sebagai masukan bagi pengklasifkasi KNN yang dioptimasi menggunakan GridSearchCV 5-fold dengan ruang pencarian 48 kombinasi parameter. Konfigurasi terbaik yang ditemukan adalah metrik jarak manhattan, K bernilai 1, dan pembobotan uniform dengan akurasi validasi silang sebesar 96.59%. Evaluasi pada data testing independen menunjukkan bahwa arsitektur hibrida VGG19-KNN mencapai akurasi pengujian akhir sebesar 97.21% dengan waktu inferensi rata-rata 1.290 milidetik per citra. Analsisi confusion matrix lintas organ membuktikan bahwa seluruh kesalahan klasifikasi bersumber dari interaksi internal antar kelas organ daun dan tidak ada satupun kesalahan yang terjadi lintas organ. Eksperimen perbandingan terhadap model pemabnding 3 kelas daun yang hanya mencapai akurasi 85.78% membuktikan bahwa kehadiran kelas organ bunga memberikan kontribusi positif terhadap kualitas representasi fitur konvolusi secara keseluruhan. Hasil tugas akhir ini menunjukkan bahwa pendekatan hibrida Transfer learning dan KNN efektif untuk klasfikasi penyakit tanaman bunga matahari dan berpotensi dikembangkan lebih lanjut sebagai sistem deteksi penyakit tanaman berbasis aplikasi di lingkungan pertanian.
=====================================================================================================================================
Sunflower (Helianthus annuus L.) is a high-value agricultural commodity that is susceptible to various diseases, including Downy mildew and Gray mold, which can cause significant yield losses if not detected at an early stage. The limited ability of farmers to visually identify plant diseases and the lack of access to agronomic experts in the field have created a growing need for an intelligent, rapid, and mobile-operable plant disease detection system. This final project aims to develop a sunflower disease classification system based on a hybrid architecture that combines VGG19 Transfer learning as a visual feature extractor with the KNearest Neighbors algorithm as the final classifier. The dataset used consists of 2,392 digital images of sunflower plants covering five condition categories, namely Downy mildew, Fresh Flower, Fresh Leaf, Gray mold, and Leaf scars, which were divided in a stratified manner at a 70:15:15 ratio into 1,674 training set images, 359 validation set images, and 359 testing set images. The VGG19 model was configured using a partial fine-tuning strategy by unfreezing the last 14 layers with a learning rate of 5e-6, achieving an optimal validation accuracy of 91.92% at epoch 15. The 512-dimensional feature vectors extracted from the GlobalAveragePooling2D layer were then used as input for the KNN classifier, which was optimized using 5-fold GridSearchCV over a search space of 48 parameter combinations. The best configuration found was the manhattan distance metric, K equal to 1, and uniform weighting, achieving a cross-validation accuracy of 96.59%. Evaluation on independent testing data showed that the VGG19-KNN hybrid architecture achieved a final testing accuracy of 97.21% with an average inference time of 1.290 milliseconds per image. Cross-organ confusion matrix analysis demonstrated that all classification errors originated from internal interactions among leaf organ classes, with no errors occurring across organ types. A comparative experiment against a 3-class leaf-only model that achieved only 85.78% accuracy confirmed that the presence of flower organ classes contributes positively to the overall quality of convolutional feature representation. The results of this study indicate that the hybrid transfer learning and KNN approach is effective for sunflower disease classification and has the potential to be further developed as a mobile application-based plant disease detection system in agricultural settings.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	VGG19, KNN, Transfer learning, Klasifikasi Penyakit Tanaman, Bunga Matahari, Ekstraksi Fitur, GridSearchCV, VGG19, KNN, Transfer learning, Plant Disease Classification, Sunflower, Feature Extraction, GridSearchCV
Subjects:	Q Science > Q Science (General) > Q337.5 Pattern recognition systems Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Ratna Amalia Azzahra
Date Deposited:	28 Jul 2026 06:16
Last Modified:	28 Jul 2026 06:18
URI:	http://repository.its.ac.id/id/eprint/138586

Actions (login required)

View Item