Enhance Support Vector Machine Models for Cassava Leaf Disease Variants Identification and Classification

Abstract

Cassava is one of the two main staple crops grown and consumed in large quantities in Nigeria's humid, sub-humid, and semi-arid regions, especially in the Southwest. Despite its usefulness and benefits to mankind, cassava is susceptible to various diseases such as “cassava mosaic disease (CMD)”, “cassava green mottle disease (CGMD)”, “cassava bacterial blight disease (CBBD)”, and “cassava brown streak disease (CBSD)”. It is very tedious and inaccurate to identify cassava disease on its plant leaf physically and even farmers in Nigeria lack good knowledge of identifying these diseases physically. However, failure to detect the disease on the plant early can affect both the quality and quantity of the product. Though, many researchers have developed different classification models based on machine learning techniques, however, many of these models are easily influenced by imbalanced datasets, the selection of irrelevant features, inappropriate selection of training and testing datasets, and fine-tuning of the hyperparameters in the classifier. This study mitigates the aforementioned problems by using equal number of datasets for the diseased dataset, 10-fold-approach was used to split the dataset into training and testing datasets, and two optimization techniques (i.e., Binary Particle Swarm Optimization (BPSO) and Reptile Search Algorithm (RSA) were separately used and also combine to fine-tune hyperparameters in the SVM and select more discriminating features to avoid overfitting.  The images of the cassava leaves were obtained from the Kaggle village datasets, scaled down and converted to grayscale pictures. Images were further enhanced and sharpened using bi-histogram equalization and morphological filtering before noise was removed using adaptive median filtering.  The Sobel edge detection method was used to separate the lesion from the uninfected parts of the leaves.  The colour moment methods and gray-level spatial dependence matrix with the haralicstatistical feature were used to extract colour and texture features, respectively. Three classification models (i.e., BPSO-SVM, RSA-SVM, and BPSO-RSA-SVM) were developed, trained and tested with the datasets.  The results show that BPSO-RSA-SVM outperforms both BPSO-SVM and RSA-SVM models in terms of all the performance evaluation metrics. Extensive experimental results show that the proposed models can extract robust features from superfluous features extracted from datasets and accurately classify cassava leaf disease.  However, this study only identified three diseases affecting cassava, and no preventative measures were proposed.