When growing plants in artificial conditions, it is important to control the lighting parameters, both natural and artificial. This study explored the feasibility of creating a low-cost portable spectroradiometer for assessing the light environment in crop production. The multi-channel spectrometer sensors AS7341 and AS7263 and the ESP32 module were selected for this task. An analysis of the problem was conducted, and ways to solve it were identified. Machine learning methods (linear regression and decision trees) were used to determine the light source type, recover the spectrum from sensor readings at individual wavelengths, and estimate the photon flux density. The obtained results were evaluated using the MAE, MRPE, and R² metrics, resulting in MRPE up to 10% for photon flux higher than 50 µmol m⁻² s⁻¹, MAE up to 10 µmol m⁻² s⁻¹ for less intensities, R² of at least 0.96 for almost all cases. It is shown that the developed algorithm achieves acceptable accuracy on various light sources, including those not used during the training process. The research results will be useful in the development of low-cost spectroradiometers for measuring illumination in crop production.

Stevia (Stevia rebaudiana Bertoni) is an important source of natural sweeteners. Since its commercial value depends on steviol glycosides, quality assessment primarily involves quantifying these compounds in leaves and shoots. While chromatography is the standard analytical method, it is labor-intensive and time-consuming; it involves multiple processing steps that may cumulatively introduce errors and remains relatively expensive. Although chromatography remains the most accurate method, this exploratory study evaluates the potential of using spectroscopy as an auxiliary method for the approximate assessment of steviol glycoside content. Leaf reflectance spectroscopy could be a simpler and more cost-effective approach. However, relationships between leaf reflectance and steviol glycoside content are indirect and mediated by physiological processes. To account for these indirect dependencies, cumulative UVB exposure was included as an additional feature because it influences both leaf optical properties and plant metabolic processes. A low-cost spectrometer was utilized as the measuring instrument. The study was conducted over a period of three months on 77 S. rebaudiana clones, divided into four groups based on their level of UVB irradiance (control without irradiation, 400, 600, and 800 μW m⁻²). Based on the collected data, linear and polynomial regression, Random Forest, XGBoost, PLSR, and ElasticNetCV models were trained. Cumulative UVB exposure was found to be the most important feature. Of the spectral features, the most informative for assessing the content of steviol glycosides were spectral indicators in the far-red and near-infrared (NIR) ranges. Our results indicate a detectable relationship, with Random Forest being the best-performing model and achieving a moderate predictive performance (R² = 0.66). Despite their limited predictive performance, the models demonstrate that leaf reflectance spectra combined with cumulative UVB exposure contain information related to steviol glycoside content. These findings support further investigation of remote sensing approaches for crop quality assessment.