Krismono Sadi, Moh Noor Al Azam

Perbandingan Performa Custom CNN, MobileNetV2, dan InceptionV3 untuk Identifikasi Penyakit Daun Tanaman

Introduction

Perbandingan performa custom cnn, mobilenetv2, dan inceptionv3 untuk identifikasi penyakit daun tanaman. Studi ini membandingkan Custom CNN, MobileNetV2, & InceptionV3 untuk identifikasi penyakit daun tanaman. MobileNetV2 unggul dengan akurasi 97.57% & lebih efisien pada batch size kecil, ideal untuk smart farming.

Abstract

Batch-size sensitivity in deep learning architectures is often assumed to behave uniformly, yet this study finds an opposite-direction pattern between architectures trained from scratch and those relying on Transfer Learning. Using a large-scale multi-commodity dataset (72,000 images, 72 disease classes from 21 commodities), Custom CNN improves consistently as batch size grows, whereas MobileNetV2 and InceptionV3 built on ImageNet pretrained weights via two-phase Feature Extraction and Fine-Tuning achieve their best performance at the smallest batch size (8), consistent with the sharp-versus-flat minima theoretical framework. The pattern is confirmed statistically using two-proportion z-tests (p<0.0001 across all Custom CNN batch-size comparisons). At its optimal configuration, MobileNetV2 significantly outperforms InceptionV3 (test accuracy 97.57% vs. 96.35%; z=4.24, p<0.0001) while being far more compact (±2.96 million parameters, roughly 8 times smaller). Explainable AI analysis (Confusion Matrix, t-SNE, and Grad-CAM) confirms this advantage stems from more separable class-level feature representations and attention concentrated on lesion regions rather than dataset artifacts, making MobileNetV2 at a small batch size a strong candidate for mobile deployment in smart farming.

Full Text

Comments

You need to be logged in to post a comment.