All projects
Machine Learning/Computer vision /Project 08 of 10 M.S. coursework

Seedling Image Classifier

A convolutional neural network that classifies plant seedling images into 12 species, with a trained model artifact included for inference and inspection.

Python Computer vision CNN Image classification Deep learning M.S. coursework

Business impact

Explores first-pass species sorting with expert review, while class-level scores show where labels are less reliable. Field-image performance was not tested.

66.62%Accuracy on 713 test images
12Species classes
0.83Best class F1

Business question

How accurately can a CNN distinguish visually similar seedling species from images?

Approach

How the work was done

  • Normalized 4,750 RGB images and used stratified splits of 3,325 training, 712 validation and 713 test images. Trained a TensorFlow/Keras CNN with augmentation and early stopping.
  • Evaluated the model on a held-out test set across 12 plant species.
  • Compared class-level F1 scores to identify where the classifier was more reliable and where visually similar species remained difficult.

Results

What the analysis shows

  • Overall test accuracy was 66.62% across the 12 classes.
  • Common Chickweed and Common wheat were the strongest classes at 0.83 F1 each.
  • Shepherd’s Purse remained challenging at 0.42 F1; Fat Hen scored 0.50 and Scentless Mayweed 0.56.

From the project

The images and the classification errors

Nine labeled photographs showing seedlings from the course image dataset.

Examples from the image dataset

These labeled examples show variation in leaf shape, lighting and background. These examples give context for the 12-species classification task.

View the source Open full size
Current saved confusion matrix for 713 held-out images across 12 seedling species.

Which species are confused

Diagonal cells are correct classifications. Off-diagonal counts show which labels need closer review, including confusion between Black-grass and Loose Silky-bent. The model correctly labeled 475 of 713 test images.

View the source Open full size

Business relevance

What this makes easier to decide

Class-specific performance suggests the classifier is best positioned as a first-pass sorting aid, with additional review for visually similar species.