iCorn: A Methodological Framework for Multi-Label Disease Classification in Maize Using Uncurated Field Images
Domain:
agricultureRecord type:
iCorn: Multi-Label Maize Disease Classification Framework
Complete implementation of iCorn—a methodological framework for multi-label
classification of maize leaf diseases from uncurated field images using deep learning,
with production-ready mobile deployment.
**Manuscript**: "iCorn: A Methodological Framework for Multi-Label Disease
Classification in Maize Using Uncurated Field Images" (Nature Scientific Reports)
**Contents**:
1. **icorn-ml-workflows** — Python training and evaluation pipeline:
- Multi-label classification training (1_multiClass.py, 2_multi_label_augmentation.py)
- Object detection preprocessing with Faster R-CNN (3_multi_label_augmentation_obj_detection_segmentation.py)
- ResNet18 backbone with PyTorch
- TensorFlow.js model export for mobile inference (pt_to_tfjs.py, 4_get_tfjs_models.py)
- Data loading and utility functions (dataset.py, utils.py)
- Requirements: Python 3.8+, PyTorch 1.x, TensorFlow 2.x, torchvision
2. **icorn-mobile-application** — React Native production app:
- Cross-platform iOS/Android mobile application
- Real-time disease classification on-device (~266–273 ms inference latency)
- TensorFlow.js model integration
- Mobile-optimized UI for farmer-facing diagnostics
- Complete build configuration and deployment scripts
**Performance Metrics** (ResNet18, multi-label classification on held-out test):
- Baseline (no preprocessing): Micro-F1 = 0.612, Macro-F1 = 0.452
- Detection-assisted (ROI preprocessing): Micro-F1 = 0.628, Macro-F1 = 0.480
- On-device latency: ~392–399 ms end-to-end (125–126 ms preprocessing + 266–273 ms inference)
**Disease Classes** (6-way classification):
Gray Leaf Spot (GLS), Northern Corn Leaf Blight (NCLB), Phaeosphaeria Leaf Spot (PLS),
Common Rust (CR), Southern Rust (SR), Other
**Note**: This repository contains code and application artifacts only. The maize disease
dataset (Craze & Berger uncurated field images, ~2,355 images) and pre-trained model
weights are not included. Refer to the Methods section of the manuscript for dataset
access and reproducibility details.