Machine learning regressor for the prediction of the SPAD value of indoor basil with RGB monitoring

Information from digital images can be automatically collected and analyzed using computer vision (CV), which allows to solve complex problems through self-learning given by artificial neural networks.The objective of this study was the evaluation of a dynamic strategy of nutrient solution management and the creation and validation of a ML algorithm to obtain a digital RGB image regressor for indoor-grown basil. The algorithm has the purpose of analyzing the images returning the prediction of the SPAD value, therefore a relative measure of the nutritional status of the plants. In commercial application, the regressor could allow fertilization to be remotely steered to avoid or remedy nutrient deficiencies. The experiment concerned the indoor growth of Ocimum Basilicum (cv. 'Genovese'), for 22 days after transplanting. During the growth cycle, five nutritional regimens were applied: surplus (160% and 130%), optimal (100%) and deficit (70% and 30%) starting from a slightly modified Hoagland nutrient solution and the experimental scheme used was randomized blocks. The architecture of the regression algorithm has foreseen the implementation of 13 layers, four validation metrics have been foreseen, such as MSE, MAE, R2 and r. The algorithm was trained for 1200 epochs with a learning rate of 0.001. Performance was encouraging, achieving low (dimensionless) MSE and MAE values of 2.92 and 1.32, respectively, and R2 and r coefficients of 0.94 and 0.97, respectively. The biomass analysis demonstrates the possibility of carrying out a dynamic strategy of nutrient management with a significant reduction in the use of fertilizers. The results obtained by the algorithm are promising, also in consideration of the discrete number of images in the dataset and the few layers, the computational lightness of which the algorithm is composed.