Image Based Automated Identification Of The Plant Species Diversity From The Western Ghats Region Using Machine Learning And CNN Based Deep Learning Techniques Using Leaf Dataset

Abstract

It becomes crucial to conserve the floral diversity of a region, as the rate at which the flora is diminishing is alarming, leading to biological degradation worldwide. Conservation is the key; to conserve one has to have a thorough knowledge of the species and also skills to identify them, which is acquired by years of rigorous training, active involvement and practice. Whereas, advancement in the field of computer vision has promoted automated plant species identification with favourable results. The paper presents the creation of a dataset with 95 plant species which are rare and endangered from the western ghats region.  The leaf Images of these plant species were captured to create the dataset. The paper further discusses the use of various image processing techniques for identification purposes. Firstly, preprocessing techniques were applied on the images to for noise reduction using thresholding techniques like Otsu’s binarization and closing morphological operations that address the problem of damaged leaves in the dataset. The shape, color, and texture features that are retrieved from the leaves for identification reasons are then explored. We used typical machine learning algorithms to achieve varying levels of accuracies. During the second phase of the experimentation, the leaf images were subjected to the process of data augmentation to synthetically expand the size and diversity of the dataset and reduce overfitting. Further, CNN models with optimised parameters were developed with fully connected layer of artificial neurons for plant species identification. This experiment demonstrated an accuracy of 84.12 in testing accuracy and 99.7% accuracy in Training accuracy.