Application of Deep Learning Technique to Real-time Crop Classification and Rice Lodging Identification

• Main Authors: Hsin-Hung Tseng, 曾信鴻
• Other Authors: 楊明德
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/66qjey

碩士 === 國立中興大學 === 土木工程學系所 === 107 === 　　At present, the investigation of agricultural disasters is time-consuming and labor-intensive. Therefore, this study develops a technology that uses the drone imagery combined with deep learning technique to study the two objectives of real-time crop classification and rice lodging identification and hopes to improve the effectiveness of agricultural disaster investigation. In the study, the area for real-time crop classification was at National Chung Hsing University, using drone collected visible images (RGB) in 2018; and the area for rice lodging identification was at Mozidunyuan, Wufeng District, Taichung City, Taiwan, using drones collected visible images in 2017 and 2019. All visible images were calculated to Excess Green (ExG) index as vegetation index, and visible information and vegetation index were used as research materials. The study applies two neural network architectures, FCN-AlexNet and SegNet for deep learning semantic segmentation architecture, and compares the speed of computation and the classifying performance on different object sizes. In the training and validating phase, the best classification result of real-time crop classification is SegNet using RGB+ExG information, the overall accuracy (OA) is 89.97%; the best result on rice lodging identification was also using SegNet with RGB+ExG, and the OA is 87.76%. Besides, the large-area rice lodging mosaic image was used to test the model of rice lodging identification, and the orthomosaic process was used to generate image tiles to facilitate the scalable adjustment during computation. The result shows that adding ExG information can let the classification be more robust. In the computation phase, the two neural network architectures are 10 to 15 times faster than the traditional Maximum Likelihood Classification (MLC), and the computation time with 40-hectare area imagery requires only about 140 seconds. In summary, this study uses drone technology to significantly improve the immediate and scientificity in agricultural surveys. Combination with RGB imagery and vegetation index in deep learning technology enhances the applicability of visible cameras in agriculture.