PORTFOLIO

CV: Cargill Agriculture Image Classifier

CNN to classify 12 plant seedlings; transfer learning and augmentation for robust performance.

Computer Vision Deep Learning TensorFlow Keras ResNet50 CNNs Data Augmentation JSON


View Notebook
  • Labeled Bar Plot of Class Distribution in Dataset
  • Sample Images from Dataset
  • Data Preprocessing Pipeline
  • Model Design / Data Augmentation / Training
  • Confusion Matrix Visualization of Model Performance
1/5
1/5
Slide 1 of 5

EXECUTIVE SUMMARY

Developed and compared multiple convolutional neural network (CNN) models to classify plant seedling images for Cargill, delivering a high-accuracy proof-of-concept for automating agricultural identification.


Goal:

Automate the classification of 12 different plant seedling species to improve efficiency and support precision agriculture.


Approach:

Built a baseline CNN and implemented a transfer learning solution with ResNet50. Applied data augmentation (random rotations) and used class weights to create a robust model that generalizes well.


Outcome:

The final transfer learning model achieved ~95% validation accuracy. This POC demonstrates a powerful and scalable solution for reducing manual labor and enhancing crop management.

THE CHALLENGE

The manual identification of plant seedlings is a time-consuming and labor-intensive process in agriculture. Cargill sought a proof-of-concept to demonstrate that computer vision could automate this task accurately and efficiently, paving the way for data-driven farming solutions.

MY APPROACH

1. Data Preparation & Augmentation:

  • Preprocessed a dataset of 4,750 seedling images across 12 species.
  • Implemented on-the-fly data augmentation to introduce rotational variance into the training set, improving model generalization.
  • Addressed a notable class imbalance by calculating and applying class weights during training.

2. Model Design & Training:

  • Baseline Model: Built a sequential CNN from scratch to establish an initial performance benchmark (~88% accuracy).
  • Advanced Model: Implemented a transfer learning strategy using a pre-trained ResNet50 model, adding custom classification layers on top.

3. Evaluation:

  • Utilized Adam optimizer and categorical cross-entropy loss.
  • Monitored learning curves and used callbacks like EarlyStopping and ReduceLROnPlateau.
  • Performed a detailed evaluation using a confusion matrix and classification report.

PERFORMANCE & VALIDATION

  • Validation Accuracy: ~95%
  • The ResNet50 model proved highly effective, demonstrating robust classification even for visually similar species.
  • Analysis confirmed the model's ability to overcome the initial dataset imbalance.

IMPACT & BUSINESS RELEVANCE

  • Accuracy: Delivers high-fidelity classification suitable for real-world agricultural applications.
  • Efficiency: Provides a foundation for an automated system that can classify crops instantly, reducing reliance on manual labor.
  • Scalability: The architecture is adaptable and can be fine-tuned for other agricultural classification tasks.

NEXT STEPS

  • Integrate with mobile or IoT camera systems for field deployment.
  • Expand dataset to include seasonal variations and multi-region crops.
  • Implement real-time inference for on-site decision making.