Rice bacterial blight and blast diseases recognition using deep learning techniques

Abstract

Rice is one of the main economic agriculture products of Thailand. One of factors which cause the decline in both quality and quantity of such product is the spreading of rice diseases which has Blight and Blast as main diseases. From the report, both diseases can be infected and destroy rice from the sapling stage to the yield stage with the damage up to 50% for Blight and 0.4% to 100% for Blast. If the diagnosis has been done accurately and in time, the damages can be diminished with the most effective solution. This research purpose is to implement the deep learning technique used specifically for recognizing rice blight and blast rice diseases images and developing the classification system for both diseases’ symptom pictures took from the rice fields. The result indicated that the YOLOv3 deep learning technique has the highest effectiveness in recognizing both diseases with the average accuracy in blight and blast rice diseases classification at 90.33% and 86.46% consecutively, while the mean average precision, mean Average Precision (mAP) is at 79.19%. Nevertheless, the reviewing for correctness YOLOv3 system by testing with a set of non-training pictures of blight and blast symptoms, the result showed that accuracy is decline to 87.50% and 74.00% respectively. Moreover, mAP is further decrease when large number of diseases added to deep learning training. The reduction of effectiveness of deep learning for disease recognition caused by the poor quality of pictures and the similarity of the disease symptom in many diseases. The successfulness in developing system using deep learning technique in rice disease recognition based on disease symptom image analysis are showed in this research. This system can be implemented as a tool for rice farmers to classify rice diseases via mobile application platform along with plant pathologist advice for more accurate diagnosis.