In closed hydroponics, fast and continuous measurement of individual nutrient concentrations is necessary to improve water- and nutrient-use efficiencies and crop production. Ion-selective electrodes (ISEs) could be one of the most attractive tools for hydroponic applications. However, signal drifts over time and interferences from other ions present in hydroponic solutions make it difficult to use the ISEs in hydroponic solutions. In this study, hybrid signal processing combining a two-point normalization (TPN) method for the effective compensation of the drifts and a back propagation artificial neural network (ANN) algorithm for the interpretation of the interferences was developed. In addition, the ANN-based approach for the prediction of Mg concentration which had no feasible ISE was conducted by interpreting the signals from a sensor array consisting of electrical conductivity (EC) and ion-selective electrodes (NO₃, K, and Ca). From the application test using 8 samples from real greenhouses, the hybrid method based on a combination of the TPN and ANN methods showed relatively low root mean square errors of 47.2, 13.2, and 18.9 mg∙L⁻¹ with coefficients of variation (CVs) below 10% for NO₃, K, and Ca, respectively, compared to those obtained by separate use of the two methods. Furthermore, the Mg prediction results with a root mean square error (RMSE) of 14.6 mg∙L⁻¹ over the range of 10–60 mg∙L⁻¹ showed potential as an approximate diagnostic tool to measure Mg in hydroponic solutions. These results demonstrate that the hybrid method can improve the accuracy and feasibility of ISEs in hydroponic applications.

Introduction

It is important to measure the growth status in real time during the cultivation of crops. In a fully controlled plant factory, the number of days after transplanting (DAT) can be an indicator of growth status due to accurate environmental management and the uniform growth of cultivated plants. This study was conducted to estimate the number of DAT of lettuce through image data using an artificial intelligence model in a closed plant factory system. The seedlings used were the 'Thimble' variety of green lettuce from Nunhems, and image data were collected according to the number of DAT. The RGB ratio was set to 8:1:1, and the amount of light was adjusted to 265(±50) μ mol•m-2 •s-1 with a photoperiod of 12 hours. The cultivation environment maintained a temperature of 19-21°C, relative humidity of 55-65%, and CO 2 in the range of 500-800 μmol•mol-1. Through three experiments conducted in identical environments, images were taken according to the date of growth from 12 specimens of 'Thimble' in total. In all three duplicated experiments, images were collected 4th, 8th, 12th, 15th, 18th, and 22nd DAT. Among collected image set, 240 images from ten objects were used for learning dataset and 48 images from two objects were used to test dataset the accuracy of the generated artificial intelligence model. The artificial intelligence model created through the convolutional neural network (CNN) generated with Python's TensorFlow had a test accuracy rate of 91.7%, and the artificial intelligence model created through a teachable machine showed accuracy of 86.7% for the test samples in this experiment. The number of DAT was predicted with the maximum probability. In this research, although the artificial intelligence model was created with a small amount of image data, it showed significantly high accuracy due to the characteristics of the plant factory where standardized image data were produced by precise environmental control and the uniform growth of the cultivated plants. Given the nature of the artificial neural network, which increase the accuracy of the model as more data are inputted, it is