Automatic lesion segmentation from rice leaf blast field images based on random forest

Rice leaf blast disease has become a threatening disease that results in severe production reduction of rice. Due to wide fields and low efficient labor, automatic image processing methods have been studied to identify crop diseases. Crop images captured in rice fields have complex background which has similar color and texture with leaf, hence it is a challenging problem to automatically segment lesion from field images. To solve this problem, we propose an automatic lesion segmentation method based on superpixel segmentation and random forest classifier. Steps of the proposed method are illustrated as follows. First of all, field images are preprocessed by resizing and color adjustment techniques. Second, preprocessed images are segmented using SLIC superpixel algorithm. Third, 32 regional features consisting of color features, shape features and texture features are extracted from each superpixel. Forth, with these features, superpixels of an image are classified by random forest classifier. When all the superpixels of an image are classified correctly, segmentation of this image is achieved. In order to evaluate the proposed method, we use 2 datasets collected at different time to test the method. Also, we evaluate superpixel classification performance and lesion segmentation performance using accuracy-sensitivity-specificity metric and Dices coefficient metric, respectively. Experimental results indicate that our proposed method achieves good segmentation performance. Therefore, the proposed method is able to segment lesion from rice images affected by leaf blast disease.