Android-Based Chili Leaf Disease Detection System Using Deep Learning For Harvest Loss Mitigation

Abstract

Bird's eye chili productivity in Indonesia faces persistent decline due to leaf disease infections. Conventional visual inspection methods by farmers show limited accuracy (65-70%) and high subjectivity, causing delayed identification and yield losses. This study develops an Android-based detection system optimized for low-end mobile devices using deep learning with MobileNetV3-Large architecture enhanced through transfer learning. The dataset contains 5,000 annotated chili leaf images across healthy leaves and three disease types (anthracnose, bacterial spot, mosaic virus). Implementation includes quantization-aware training and TensorFlow Lite conversion for mobile optimization. Model evaluation uses 5-fold cross-validation with accuracy, precision, recall, and F1-score metrics. The model achieved 94.34% classification accuracy with 94.5% precision. Quantization reduced model size by 92.2% (75.6 MB to 5.9 MB) with only 0.3% accuracy loss. The Android application operates in real-time on 3GB RAM devices with inference latencies below 100ms. This system provides an effective solution combining high accuracy with computational efficiency for early chili leaf disease detection, supporting sustainable farming in Indonesia.

Index Terms---Deep learning; Mobile application; Plant disease detection; Chili leaf;