Handling hypercolumn deep features in machine learning for rice leaf disease classification

Abstract

Rice leaf disease, which is a plant disease, causes a decrease in rice production and more importantly, environmental pollution. 10–15% of the losses in rice production are due to rice plant diseases. Automatic recognition of rice leaf disease by computer-assisted expert systems is a promising solution to overcome this problem and to bear the shortage of field experts in this field. Many studies have been conducted using features extracted from deep learning architectures, so far. This study includes keypoint detection on the image, hypercolumn deep feature extraction from CNN layers, and classification stages. The hypercolumn is a vector that contains the activations of all CNN layers for a pixel. Keypoints are prominent points in the images that define what stands out in the image. The first step of the model proposed in this study includes the detection of keypoints on the image and then the extraction of hypercolumn features based on the interest points. In the second step, machine learning experiments are carried out by running classifier algorithms on the features extracted. The evaluation results show that the proposed approach in this paper can detect rice leaf diseases. Furthermore, the Random Forest classifier presented a very successful performance on hypercolumn deep features with 93.06% accuracy, 89.58% sensitivity, 94.79% specificity, and 89.58% precision. As a result, the proposed approach can be integrated into computer-aided rice leaf disease diagnosis systems and so support field experts.