seedling-classification-computer-vision

Context

Recently, the agricultural sector has faced a pressing need for modernization due to the extensive manual labor required to monitor plant growth accurately. Although agricultural technology has advanced, workers still spend considerable time and effort distinguishing between different plants and weeds. There's a significant opportunity for this trillion-dollar industry to benefit from innovations that reduce manual labor. Artificial Intelligence stands out as a key technology that can transform this field by dramatically decreasing the time and effort needed to identify plant seedlings using AI and Deep Learning. This enhancement not only makes the process more efficient than traditional methods but also improves crop yields. It allows workers to focus on higher-level decision-making in agriculture and promotes more sustainable environmental practices over the long term.

Objective

The goal of this project is to develop a Convolutional Neural Network that can accurately classify plant seedlings into their specific categories. To ensure our code runs efficiently, we'll utilize the GPU runtime available on Google Colab.

Download Data File (images.npy)

# Ensure to comment out the 'upload file' code prior to executing the script
$ python data/bucket-storage.py

Data dictionary

The Signal Processing group at Aarhus University, together with the University of Southern Denmark, has recently made available a dataset that includes images of unique plants from 12 different species.

• The data files are named:

  • images.npy
  • Labels.csv

• To manage the large volume of data efficiently, the images have been compiled into the images.npy file, and the labels are stored in Labels.csv.

• This setup allows me to work on the project seamlessly without concerns about data size.

• The project's objective is to develop a classifier that can accurately identify the species of a plant based on its image.

List of Species

• Black-grass
• Charlock
• Cleavers
• Common Chickweed
• Common Wheat
• Fat Hen
• Loose Silky-bent
• Maize
• Scentless Mayweed
• Shepherds Purse
• Small-flowered Cranesbill
• Sugar beet

Skills

• Exploratory data analysis
• Data preprocessing
• Data preparation for modelling
• Model building (CNN)
• Performance evaluation
• Final model selection

Actionable Insights and Reccomendations

Model Performance

• From the confusion matrix across all models, it's clear that Model V1 outperformed the others by accurately predicting most of the classes.
• Model V1 achieved a test accuracy of 78%.
• Right from the initial iteration, this model was capable of correctly predicting 8 out of the 10 plant seedling categories.
• Techniques such as batch normalization and data augmentation did not enhance the performance of the model.
• Additional adjustments and fine-tuning may be required to effectively utilize these techniques and improve the model further.

Recommendations to Further Enhance the Model

1. Experiment with Filter Variations : Experimenting with different filter sizes and the number of filters can help in capturing a variety of features from the images, which might improve the model’s ability to generalize across different plant seedling types.
2. Utilize Original Image Sizes : Training the models on the original image size of 128x128 pixels, instead of reducing them to 64x64, could preserve important details that are crucial for accurate classification.
3. Increase Samples for Underrepresented Classes : Specifically increasing the number of samples for classes 0 and 4 could help improve the model's accuracy, especially if these classes are currently underrepresented and affecting the model's learning.
4. Refine Data Augmentation and Dropout Techniques : Enhancing the data augmentation strategies and adjusting the dropout rates could potentially increase the robustness of the model, preventing overfitting and promoting better generalization.
5. Explore Transfer Learning : Implementing other transfer learning architectures can leverage pre-trained networks which have proven successful in similar tasks, potentially boosting performance and efficiency in training.