Agriculture Leaf Vision

Agriculture Leaf Vision is a lightweight computer vision API that identifies plant species (or diseases) based on uploaded leaf images. It's built using transfer learning with CNN EfficientNetB2. I trained it on a PlantVillage dataset, and deployed to a serverless cloud backend.

The project was developed to deepen my understanding of image preprocessing, transfer learning, and end-to-end ML deployment with minimal cloud cost. Originally deployed to Vertex the costs we to great for a side project.

---

Try the App

Live Demo:
https://agricultural-leaf-vision-7fab2.web.app/

You can upload a leaf image and receive a class prediction from the live API.
⚠️ Note: The model is not trained to production quality — it often predicts the same label ("Pepper") for all inputs. It is intentionally undertrained to demonstrate infrastructure, not accuracy.

---

What It Does

• Takes an uploaded image (leaf)
• Preprocesses the image to the correct input size
• Applies data augmentation during training
• Uses a pretrained CNN backbone (MobileNetV2 or EfficientNet)
• Fine-tunes the model on a small labeled dataset (PlantVillage)
• Returns class prediction via a REST API

---

Machine Learning Details

Type	Value
ML Category	Supervised Learning
Model Type	Convolutional Neural Network
Strategy	Transfer Learning
Backbone Models	EfficientNetB0
Pretrained On	ImageNet
Fine-Tuned On	PlantVillage (plant disease)
Data Augmentation	Rotation, flip, zoom, contrast
Loss Function	Categorical crossentropy

---

Model Evaluation

While the training curves suggest promising performance, the confusion matrix reveals signs of overfitting — the model learned the training data well but fails to generalize to unseen examples.

Accuracy Curve

Loss Curve

Confusion Matrix

🧠 Note: The model shows high training accuracy and low loss, but the confusion matrix highlights that the classifier consistently mispredicts most classes — suggesting the model is overfitting to a narrow subset of the training distribution.

---

🧪 Preprocessing & Training Pipeline

To adapt a generic vision model to agricultural data I applied these techniques:

• Image resizing and normalization to match backbone input specs
• Augmentation: horizontal/vertical flips, zoom, rotation, contrast variation
• Batching with stratified shuffling
• Used early stopping and learning rate scheduling

Full training from scratch was avoided. Transfer learning was sufficient given dataset size.

---

Tech Stack

• Python 3.10
• TensorFlow / Keras
• tf.keras.applications (MobileNetV2, EfficientNetB0)
• Docker (containerizing the API)
• FastAPI (optional REST layer)
• Google Cloud Run (serverless model hosting)
• Artifact Registry (image storage)
• Firebase (optional demo frontend)
• Vite (for future frontend deployment)

---

Lessons Learned

• Observed how dataset quality directly impacts model output
• Evaluated tradeoffs between model size and accuracy
• transfer learning in a resource-constrained setting
• Built a Dockerized ML inference API
• Deployed with Cloud Run using minimal budget
• Fine-tuned image classification with augmented data
• Explored tradeoffs between accuracy vs model size
• Realized importance of input resolution and normalization for pre-trained CNNs

---

Known Limitations

• The model is not accurate — due to minimal training and limited dataset coverage
• It frequently predicts the same class ("Pepper") regardless of input
• Dataset coverage is limited to a few categories and model
• This project is for educational/demo purposes only and not suitable for real agricultural use
• Best results come from clear, close-up leaf images

---

📚 Dataset

• PlantVillage Dataset
  Kaggle Link

---