Development of the Automatic Control Systems for the Centella Asiatica (L.) Urb. Greenhouse by Using Internet of Things (IoT) and the Classification of Centella Asiatica (L.) Urb. by Using Deep Learning

Abstract

The objectives of this research were divided into two parts: 1) part 1 was to develop the automatic control systems for the Centella asiatica (L.) Urb. greenhouse consisting of sensors used to check temperature and soil humidity and a N P K element sensor by sending information to a NetPie sever working together with the mages from the CCTV cameras in the greenhouse and the AI server to control the work of the water reduction system and the misting system, the greenhouse being able to be controlled at the greenhouse, through the internet system, and through the automatic artificial intelligence  systems and 2) part 2 was to study the classification of the small, medium, and large sizes of the Centella asiatica (L.) Urb. images by using the deep learning to develop an artificial neural structure classification model, Convolutional Neural Network (CNN), by using the models of VGG16, ResNet50, and InceptionV3 coupled with transfer learning using data sets collected by the researchers and from the internet.

The research findings were as follows.

1. The experts and the users in a total of 10 people were satisfied by using the automatic control systems for the Centella asiatica (L.) Urb.greenhouse at a high level ( = 4.16, S.D. = 0.74) and according to the evaluation of the automatic artificial intelligence systems, the VGG16 model was at the highest efficiency with the accurate value of 0.82, followed by the InceptionV3 model with the accurate value of .73 and the ResNet50 model with the accurate value of .60.  The InceptionV3 model and the ResNet50 model were at a good efficient level.  
2. 300 Centella asiatica (L.) Urb.images were tested and they were classified in the small sizes correctly 83 images and incorrectly 19 images, in the medium sizes, 82 images were classified correctly and 18 images were classified incorrectly, and in the large sizes, 80 images were classified correctly and 18 images were classified incorrectly.