Plant Toxicity Classification by Image

Since differentiating between dangerous and safe plants is a complex task for a human brain, this study approaches the issue through machine learning models starting with a convolutional neural network (CNN) and discovering that a logistic regression model—trained on a dataset with manually designed features—has the best performance with the particular dataset used.

Poisonous and non-poisonous plants have extremely similar visible features to any non-botanist, which puts those in danger who are frequently present in areas containing various plants. Failure to distinguish harmful plants from safe ones puts several people at a high risk of accidents and potential health issues after contact with a toxic plant. While previous work has found ways to classify specific types of plants, a limited amount of research has been done with toxic and non-toxic plants of several species. Since differentiating between dangerous and safe plants is a complex task for a human brain, this study approaches the issue through machine learning models starting with a convolutional neural network (CNN) and discovering that a logistic regression model—trained on a dataset with manually designed features—has the best performance with the particular dataset used. The neural network demonstrated overfitting which was likely caused by the inconsistent backgrounds of images within the dataset. The logistic regression model achieved an accuracy of 97.37% in its predicted classifications for the plants. The best-performing logistic regression model contained the three-leaf and dark red stem features indicating that these two features are the most reliable ones used to distinguish between toxic and non-toxic plants.

Explore More!

Research Paper