Knowledge transfer for adapting pre-trained deep neural models to predict different greenhouse environments based on a low quantity of data

Abstract

Deep learning is the state-of-the-art application of machine learning in many fields, and this technology has also been applied in agriculture. A large quantity of data needs to be provided to the deep learning models in the training procedure; however, sufficient data may not be provided when considering agriculture applications. Transfer learning, which is a learning strategy for rapid and easy adaptation of a pre-trained model, can be a solution for limited agricultural data. Therefore, the objective of this study is to verify the adaptability of a pre-trained model that predicts the environmental variables of a greenhouse by retraining the model with data from a new cultivation condition, using the transfer learning technique. As a result, the transfer learning methodology was applied to five common deep learning models. Twenty-seven greenhouses (14 sweet peppers and 13 tomato cultivations) in various regions of South Korea provided the experimental dataset to this research. The analyzed environmental variables are the internal temperature, relative humidity, radiation, CO2 concentration, and external temperature. Before the transfer learning procedure is conducted, some layers from pre-trained models were replaced with new layers. The model was, thereafter, re-trained with a new test dataset. The best model in the training procedure was BiLSTM, resulting in an average R2 of 0.69. The models could predict the tendencies of the environmental changes, indicating that they were adequately trained. The most accurate deep-learning model considering the transfer dataset was the transferred BiLSTM, with an average R2 of 0.78 and 0.81 for sweet pepper and tomato datasets, respectively. The accuracies of most transferred models are higher than those of the corresponding deep-learning models. As a result, transfer learning can be used to adapt previously trained deep-learning models, enabling them to predict the microclimates of a greenhouse with scarce data. Furthermore, advanced transfer learning strategies would increase the performance of the transferred models analyzed in this study.