Effects of Partial Dependency of Features and Feature Selection Procedure Over the Plant Leaf Image Classification

Abstract

The process of taxonomic classification of plant species has been carried out by the botanist since centuries, by observing their roots, shoots and flowers. It is the age of modernization; roads, buildings and bridges are fast replacing the vegetation, even before the botanist might personally get a chance to look at them. Therefore, the role of computer vision is justified for fast classification of plant species before they become extinct. The sole purpose of this research work is to increase the predictive classification accuracy of plant species by using their shape and texture features obtained from the digital leaf images of dorsal sides. Since the geometrical shape features of the leaves alone are not able to provide better predictive classification accuracy results, therefore, the texture features have been clubbed together to achieve higher order of accuracy results. This leads to increase in the data size. Therefore, in order to reduce the feature dataset, random feature selection procedure has been adopted, which selects features on the basis of weights of attributes. The justifiability of the features selected has been carried out by using the feature importance plots and strengthened further by the partial dependency plots having been drawn to see their final inclusion into the final feature selection dataset. The results exhibited by the shape, shape subset, texture, texture subset of feature dataset as well as the combined dataset are quite exemplary and worth showcasing. In spite of the fact that the geometrical shapes of many of the leaves may be the same or almost same, the combined shape and texture features can be a suitable alternative for improvement in predictive accuracy results.