Feature Selection Using Graph-Based Clustering for Rice Disease Prediction

Abstract

Rice is one of the main crop cultivated all over the world. It is attacked by several diseases and pests that reduce quantity and degrades quality of the product. There are different types of rice diseases like Leaf Blast, Brown Spot, Stem Rot, Bakanae, yellow Dwarf which damage various parts of the plant. Disease identification at the early stage and taking precautions in time help the farmers to sustain both the quality and quantity of the product. The present work extracts different types of features from the disease portions (i.e., images) of the plants and identifies the most valuable features that can distinguish the disease types. To identify the most valuable features, initially, a weighted graph is constructed with extracted features as nodes and similarity between every pair of features as the weight of the corresponding edge. Based on the weights assigned to the edges, importance of each node of the graph is calculated. Finally, a graph-based clustering algorithm namely, Infomap clustering algorithm is applied on the graph to partition it into a set of connected subgraphs. Then the most influential node from each subgraph of the partition is selected and this subset of nodes is considered as important feature subset useful for rice disease prediction. The experimental result shows the effectiveness of the proposed method.