Evaluation of Homogeneous and Heterogeneous Distributed Ensemble Feature Selection Approaches for Classification of Rice Plant Diseases

摘要

Nowadays, agricultural products have very few nutritional values due to various factors, such as environmental issues, genetic dilution effect, and excessive use of chemical pesticides to cure plant diseases. Currently, India, as well as the whole world, is encountering nutritional starvation. In developing countries, pesticide poisoning deaths are a significant problem nowadays. To reduce the spread of diseases and to facilitate an effective management strategy, developing a computer-aided diagnosis system for early recognition and classification of rice plant diseases is essential. Technological innovations in the machine learning models for predicting plant diseases using classifiers and discoveries in digital image processing techniques play a significant role in the agriculture industry. However, the accuracy of these classifiers purely depends on training the network with relevant features. Training these models with relevant features avoids overfitting and underfitting problems. Hence Feature Selection (FS) task has turned out to be an essential pre-processing task prior to classification, which not only reduces the computational cost of training and memory requirement; but also reduces the generalization capability of the network due to the “curse of dimensionality” issue. Therefore, to address this issue, a proper FS technique is required for Dimensionality Reduction (DR). This paper aims to evaluate the Ensemble Feature Selection (EFS) using homogeneous and heterogeneous approaches to classify rice plant diseases. ReliefF, Correlation-Based Feature Selection (CFS), Pearson Correlation Coefficient (PCC), Mean-Variance (MV) are the FS algorithms used for Heterogeneous Distributed Ensemble Approach (HeDEA), and an aggregator is used to select the final subset of features. ReliefF FS algorithm is used in Homogeneous Distributed Ensemble Approach (HoDEA). Deep Convolution Neural Network (DCNN) is used as a classifier for both approaches. HeDEA offers better performance when compared to the performance of the individual FS method.