Deep learning models based on hyperspectral data and time-series phenotypes for predicting quality attributes in lettuces under water stress

Abstract

Efficiently analyzing the relationship between plant phenotypes, quality, and resistance remains challenging. In this study, deep learning models based on hyperspectral data and time-series phenotypes from the high-throughput plant phenotyping (HTPP) platform were proposed to predict quality attributes of lettuce under water stress, including SSC, pH value, nitrate (NO3–), and calcium (Ca2+). First, deep learning models were developed using the Inception module and raw hyperspectral data to non-destructively predict the above quality attributes. In addition, partial least squares regression (PLSR) and support vector regression (SVR) were used to develop prediction models to evaluate performance of the Inception module. Second, the residual and attention modules were implemented to enhance performance of the Inception module. Third, time-series phenotypes were fed into four recurrent neural networks (RNNs), such as TimeDistributed (TD), long short-term memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional RNN (BRNN) and combined with the optimal deep learning models based on hyperspectral data to enhance prediction precision. The optimal performance of the Inception-residual-attention-TD model was achieved with Rp2 of 0.8900 and 0.9435 for SSC and NO3–, respectively. The Inception-residual-TD model with Rp2 of 0.9583 provided the most accurate pH value prediction. With Rp2 of 0.8716, the Inception-attention-LSTM model provided the most accurate prediction of Ca2+. Meanwhile, the Inception-residual-TD model was used to detect water stress, producing an Accuracyp of 98.86%. The Inception-residual model based on pixel-wise hyperspectral data was used to visualize the spatial distribution of pH value, and the distribution map was used to detect early water stress. The results indicate that deep learning models can use hyperspectral data and time-series phenotypes to predict lettuce quality attributes and water stress in a non-destructive manner.