Development of a smart machine-vision-based system to detect water stress in greenhouse tomato plants

Abstract

Timely detection of water stress in agricultural crops is important. In this paper, a smart classification algorithm was developed to detect water stress in tomato plants that were grown in the greenhouse. During the growth period, thermal and visible light images were acquired from the canopy tops in two states: (1) plants in normal conditions; and (2) plants under water stress. Images were obtained using a camera that recorded simultaneous frames of thermal and visible (red, green, and blue (RGB)) features. Based on these features, 22 parameters were defined and applied to classify the image frames. In order to develop an efficient algorithm, principal component analysis (PCA) was applied to optimize the classifying of parameters. For normalizing the data in PCA, 6 normalization methods were applied and assessed. Among them, peak normalization was the best as its PC1 and PC2 described 94% and 5% of total variation, respectively. Based on the PCA results, 9 parameters were found with most loadings as the most effective indexes that all obtained from the visible features. In other words, the thermal features were not as useful for detecting plant water stress. These parameters were used in multilayer perceptron neural networks (MLPNN) to develop the classification algorithm. The resulting mean-square error and r values for the MLPNN with ten hidden layer were 6.05×10^-3 and 0.9905, respectively which shows the robustness of the classification algorithm. This algorithm accuracy was 83.3%.