Analisis Perbandingan Kinerja Model Transfer Learning VGG16 Dan CNN Baseline Dalam Deteksi Penyakit Tanaman Tomat Berdasarkan Citra Daun

Budi Riansyah (Universitas Teknokrat Indonesia)
Styawati Styawati (Universitas Teknokrat Indonesia)

Dinamik
Vol 31 No 2 (2026)

Tomatoes (Solanum lycopersicum) are a vital horticultural commodity that is highly susceptible to pathogen attacks, making visual symptoms on leaves the main indicator for early detection. However, these automatic detection efforts face significant challenges related to the limited variety and number of image datasets, which often hinder the performance of Deep Learning models. This study aims to compare the performance of Baseline CNN (training from scratch) with VGG16 (fixed feature extraction) on 10 classes of tomato leaf diseases. The evaluation results show that Baseline CNN achieved an accuracy of 87.30% and VGG16 achieved 85.30%. The advantage of the Baseline model lies in its flexibility in learning visual features from scratch, making it more adept at capturing specific details such as ring patterns in Early Blight. In contrast, VGG16 provides computational efficiency with 40% less parameter training load and proves superior in recognizing texture patterns, such as in Bacterial Spot (Recall 97%). In conclusion, although Transfer Learning is efficient and robust, this model has limitations in understanding the unique characteristics of plants (semantic gap).