Early detection of drought stress in tomato from spectroscopic data: A novel convolutional neural network with feature selection

Abstract

The yield and quality of tomato (Solanum lycopersicum L.) crops are lower when the plants are exposed to drought stress. Drought stress can be prevented through timely irrigation if it is identified early. Thus, this study modified the one-dimensional spectrogram power net (1D-SP-Net) to formulate a 1D convolutional neural network with an embedded residual global context (ResGC) block; this network, called 1D-ResGC-Net, processes visible and nearinfrared (Vis/NIR) spectroscopy data of tomato leaves to identify the early signs of drought stress. In evaluation experiments, the proposed 1D-ResGC-Net model outperformed partial least squares discriminant analysis (PLSDA) and random forest (RF) models. Gradient-weighted class activation mapping, variable importance in projection, and variable importance were used to identify the most important feature bands (i.e., those that were most strongly associated with drought stress) as output by the 1D-ResGC-Net, PLSDA, and RF, respectively. The 1D-ResGC-Net model achieved 90% accuracy when the 15 most important feature bands were used; by contrast, the PLSDA and RF models required more than 90 of the most important feature bands to reach 90% accuracy. When the number of input features is similar, the accuracy of 1D-SP-Net and 1D-ResGC-Net is very close. However, when the number of input features is reduced, the accuracy of 1D-SP-Net will be much lower than 1DResGC-Net. In summary, 1D-ResGC-Net offers greater accuracy at a lower cost.