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

Highlights

• Deep learning models need big data for training, but data is not enough in agriculture.

• Transfer learning was applied to analyze the greenhouse environment.

• Most of the transferred models showed good performances.

• The trained models can predict the microclimate of a greenhouse with scarce 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.

Introduction

In agriculture, the amount of stored data is increasing, and the types of methodologies for data analyses are diversifying (Muangprathub et al., 2019, Sarker et al., 2019). Wireless networks with IoT applications are the main cause for the increase in data accumulation (Kochhar and Kumar, 2019). Novel data types, such as hyperspectral imaging, have also been used in previous research for yield prediction (Hassanzadeh et al., 2020). Owing to the improvement of sensors and computational performance, collecting and monitoring multidimensional data have continuously been studied. (Wolfert et al., 2017, Khanna and Kaur, 2019).

With an expectation to understand a phenomenon from a stable database, many types of methodologies were adapted in the agriculture area. As the deep learning methodology represents the state-of-the-art in many areas owing to its performance, it has also been used to solve problems in the agriculture, such as fruit segmentation, autonomous greenhouse control, and estimation of root-zone ion concentration (Barth et al., 2019, Hemming et al., 2019, Moon et al., 2019a). Deep learning became prevalent, because the adaptation of this methodology to address problems was relatively simple.

When analyzing phenomena in greenhouses, active control of the internal environment improves crop quality and yield. The active control of the environment surmounts the regional limitations of plant growth (Sethi et al., 2013, Shamshiri and Ismail, 2013). As a result, optimal control of the internal environment is considered an important goal regarding greenhouse cultivation. The research of new technologies or approaches that led to optimizations have been conducted (Van Beveren et al., 2015, Atia and El-madany, 2017, Tsai et al., 2020).

One of the new technologies, deep learning, requires a large amount of data (Sun et al., 2017). However, in a greenhouse environment, the vulnerability of sensors to relative humidity and temperature, which may result in data loss, and the restricted number of cultivated crops affect the data collection (Moon et al., 2019b, Lee et al., 2020). Moreover, disparate tendencies of plant environment and growth due to target crop, season, and cultivation strategy lead to the (Serret et al., 2018) requirement of more diverse data (Jones et al., 1991, Van Henten and Van Straten, 1994, Wubs et al., 2012). As a result, the collection of sufficient data is a difficult task in this environment, resulting in lower performance of deep learning models when compared to that in other fields, where data are abundant (Kamilaris and Prenafeta-Boldú, 2018).

Owing to the restricted data conditions in agriculture, better adaptations of the deep learning models are necessary. Transfer learning, which is one of the machine learning studies, may be an alternative solution to this problem, as it is applied for situations of limited available data. In the machine learning field, derived models need to learn tasks without prior knowledge of the target task. In the transfer learning methodology, the knowledge derived from data can be transferred from some previous tasks (Pan and Yang, 2009).

As a result, pre-trained models have been actively adopted, by varying the structure of the network, and, afterward, retrained with a new dataset (Shin et al., 2016, Gómez-Valverde et al., 2019, Thenmozhi and Reddy, 2019). Moreover, this methodology has improved the performance of the model (Devlin et al., 2018). Therefore, if the pre-trained plant environment can be adapted to a new cultivation condition, deep learning model can show better performance. This study aimed to verify the adaptability of a pre-trained model of a greenhouse environment by retraining it using transfer learning with a new cultivation condition.