Automatic identification of pepper seedling using Mask R-CNN for real time stress monitoring in the plant factory Sumaiya Islam (a), Md Nasim Reza (a, b), Shafik Kiraga (a), Shahriar Ahmed (b), Dong-Hee Noh (c), Sun-Ok Chung (a, b*)
(a) Department of Smart Agricultural Systems, Graduate School, Chungnam National University, Daejeon 34134, Republic of Korea
(b) Department of Agricultural Machinery Engineering, Graduate School, Chungnam National University, Daejeon 34134, Republic of Korea
(c) Jeonbuk Regional Branch, Korea Electronics Technology Institute (KETI), Jeonju 54853, Republic of Korea
Pepper seedlings need nutrients to grow and develop. Preventing nutritional stress in pepper seedlings at the early stages of development and avoiding yield loss need careful monitoring of their progress. Using only a visual examination might lead to an incorrect diagnosis, preventing seedlings that have been injured from receiving timely treatment. In this work, the goal was to identify the nutritional stress state of pepper seedlings using the enhanced Mask R-CNN, and to forecast the amount of stress using the suggested approach. An automated image capture system was used to take images of pepper seedlings from the plant factory. In the images, the seedlings are between one and three weeks old. Resnet-101 architecture with transfer learning and fine tuning were applied in the experiment to construct the enhanced Mask R-CNN structure. We proceeded with the training of the model by using images of pepper seedlings. There are 150 photos included in the dataset that are used for the training phase, and there are 40 images used for testing. In addition, image augmentation was performed in order to expand the variety of the dataset. The results of the experiments indicated that the suggested method obtained the highest level of accuracy of 91%, and 83%, respectively for the training and test datasets. Our proposed model delivered equivalent performance to comparable deep learning architectures despite having far less trainable parameters. In our future study, we will extend our CNN-based model to monitor different seedlings image real time with greater precision.
