Spatiotemporal Deep Learning Model for Prediction of Taif Rose Phenotyping

Despite being an important economic component of Taif region and the Kingdom of Saudi Arabia (KSA) as a whole, Taif rose experiences challenges because of uncontrolled conditions. In this study, we developed a phenotyping prediction model using deep learning (DL) that used simple and accurate methods to obtain and analyze data collected from ten rose farms. To maintain broad applicability and minimize computational complexity, our model utilizes a complementary learning approach in which both spatial and temporal instances of each dataset are processed simultaneously using three state-of-the-art deep neural networks: (1) convolutional neural network (CNN) to treat the image, (2) long short-term memory (LSTM) to treat the timeseries and (3) fully connected multilayer perceptions (MLPs)to obtain the phenotypes. As a result, this approach not only consolidates the knowledge gained from processing the same data from different perspectives, but it also leverages on the predictability of the model under incomplete or noisy datasets. An extensive evaluation of the validity of the proposed model has been conducted by comparing its outcomes with comprehensive phenotyping measurements taken from real farms. This evaluation demonstrates the ability of the proposed model to achieve zero mean absolute percentage error (MAPE) and mean square percentage error (MSPE) within a small number of epochs and under different training to testing schemes.

Automated summary

In this study, a deep learning-based phenotyping prediction model for roses was developed, utilizing three different deep neural networks to process images, time series, and obtain phenotypes. The model shows good predictability with incomplete or noisy datasets, and its effectiveness has been validated through comparison with real farm data.