Transfer learning for crop disease detection using smartphone imagery in precision agriculture

Abstract

This paper demonstrates transfer learning from ImageNet-pretrained EfficientNet-B4 for detecting 26 crop diseases across wheat, rice, maize, and potato from smartphone-captured leaf images. A custom data augmentation pipeline simulates varying lighting and camera conditions. The mobile-optimized model achieves 95.8% top-1 accuracy on a field-collected dataset of 54,000 images and runs at 15 FPS on mid-range Android devices using TensorFlow Lite.

Topics

Full Text Preview

The dataset comprises 54,000 leaf images captured by farmers using smartphones (Redmi Note 10, Samsung A52, Pixel 4a) across 8 Indian states over two crop seasons (Kharif 2023, Rabi 2023-24). Images are labeled into 26 disease classes and 4 healthy classes. EfficientNet-B4 pretrained on ImageNet is fine-tuned with progressive resizing (224 to 380 pixels). The TFLite model (23 MB) achieves 15 FPS inference on a Snapdragon 695 chipset. The Android app provides disease identification and treatment recommendations in Hindi and English.

Published Through