Application of electronic nose and machine learning used to detect soybean gases under water stress and variability throughout the daytime.

The development of non-invasive methods and accessible tools for application to plant phenotyping is considered a breakthrough. This work presents the preliminary results using an electronic nose (E-Nose) and machine learning (ML) as affordable tools. An E-Nose is an electronic system used for smell global analysis, which emulates the human nose structure. The soybean (Glycine Max) was used to conduct this experiment under water stress. Commercial E-Nose was used, and a chamber was designed and built to conduct the measurement of the gas sample from the soybean. This experiment was conducted for 22 days, observing the stages of plant growth during this period. This chamber is embedded with relative humidity [RH (%)], temperature (°C), and CO concentration (ppm) sensors, as well as the natural light intensity, which was monitored. These systems allowed intermittent monitoring of each parameter to create a database. The soil used was the red-yellow dystrophic type and was covered to avoid evapotranspiration effects. The measurement with the electronic nose was done daily, during the morning and afternoon, and in two phenological situations of the plant (with the healthful soy irrigated with deionized water and underwater stress) until the growth V5 stage to obtain the plant gases emissions. Data mining techniques were used, through the software "Weka™" and the decision tree strategy. From the evaluation of the sensors database, a dynamic variation of plant respiration pattern was observed, with the two distinct behaviors observed in the morning (~9:30 am) and afternoon (3:30 pm). With the initial results obtained with the E-Nose signals and ML, it was possible to distinguish the two situations, i.e., the irrigated plant standard and underwater stress, the influence of the two periods of daylight, and influence of temporal variability of the weather. As a result of this investigation, a classifier was developed that, through a non-invasive analysis of gas samples, can accurately determine the absence of water in soybean plants with a rate of 94.4% accuracy. Future investigations should be carried out under controlled conditions that enable early detection of the stress level.