AI-Powered Apple Leaf Specialist

This project is a part of the Machine Learning Operations (AAI-540-02) course in the Applied Artificial Intelligence Master Program at the University of San Diego (USD).

-- Project Status: Completed

Introduction

AI-Powered Apple Leaf Specialist helps growers and hobbyists understand what’s going on when apple leaves show spots, rust, or other symptoms. The app looks at a few photos, predicts the likely condition, and gives clear, safety-first steps the user can take.

• Model objective: classify an apple leaf photo into a small set of conditions and return a short action plan.
• ML problem type: supervised multi-class image classification with an unsure path for low-confidence cases.

Objectives

• Provide a condition label and a confidence score for each uploaded photo.
• Offer plain-language guidance for each condition, plus a prevention tip.
• Keep costs low by training in Colab and deploying a short demo on SageMaker (serverless).
• Start with PlantVillage (Apple subset) and, if resources allow, add Plant Pathology 2020 and PlantDoc later.

Dataset

This project uses the PlantVillage (Apple) subset, a research dataset of labeled apple leaf images for disease recognition. Each image is a single leaf with a class label. For apple, the core classes are:

• healthy
• scab (Venturia inaequalis)
• rust (cedar apple rust)
• black_rot (Botryosphaeria obtusa)

Why this dataset

• Consistent images (single leaf, plain backgrounds) make ingestion and training straightforward.
• Labels map cleanly to a small, meaningful set of conditions.
• Available via TensorFlow Datasets (TFDS), which keeps the pipeline reproducible.

What to expect

• Clean backgrounds help the model learn disease patterns quickly.
• Less representative of “in-the-wild” photos; strong augmentations and careful validation help address this.
• No personal data is expected (plant leaves, not people).

How it's used here

• Serves as the sole source for the baseline model in this project.
• Labels are aligned to a canonical set: healthy, scab, rust, black_rot (with an unknown route at inference time for low confidence).
• Images are standardized to a fixed size and split 70/15/15 (train/val/test)

Solution Overview

Methods Used

• Data wrangling and label harmonization (Colab)
• Exploratory Data Analysis (EDA) & data sanity checks
• Class imbalance handling (class-aware augmentation; optional focal loss/weights)
• Image augmentation (brightness/contrast jitter, blur/noise, crop, perspective, MixUp)
• Transfer learning (ResNet-18) and probability calibration (temperature scaling)
• Hyperparameter tuning (seeded grid) and threshold selection (τ=0.65)
• Error analysis with confusion matrices & per-class metrics
• Batch inference for MLOps smoke tests (n=20 cost-bounded)

Technologies

• Python, PyTorch/TorchVision
• scikit-learn (metrics, calibration helpers)
• Google Colab (wrangling/EDA/augmentation)
• AWS SageMaker (training jobs, Model Registry, Batch Transform)
• Amazon CloudWatch (infra dashboards & alarms)
• SageMaker Clarify (bias/fairness reports)
• Amazon S3 (datasets, manifests, reports)
• ONNX / TorchScript (export); Streamlit (app)

Results

• Test accuracy ~0.989, macro-F1 ~0.986; strong per-class precision/recall.
• Batch smoke test (n=20) shows high confidence, with unknown routing for low-confidence cases.
• Clarify shows no major slice-wise disparity across brightness_bin; findings captured in the model card.
• CloudWatch dashboard confirms low p90 latency and zero 5XX during dev runs.
• Performance is consistently strong across classes; rust has slightly lower precision but perfect recall.
• Results align with validation trends—no signs of overfitting.
• Ready for packaging and deployment; consider targeted augmentation or threshold tuning if rust precision needs a bump.

License

This project is licensed under the MIT License. See the LICENSE file for more details.

Acknowledgments

Thanks to Professor Jules Malin for guidance and feedback. Credit to the PlantVillage team and the Plant Pathology 2020 organizers for making data available for research and education.