Modelos de aprendizado de máquina para detectar estresse abiótico e clorofila total em folhas de trigo usando espectroscopia óptica

Abstract

The detection of abiotic stress and the estimation of chlorophyll are fundamental for determining plant health. There are different methods to estimate and detect these values in plants using optical spectroscopy. The methods can be destructive (absorption spectroscopy) or non-destructive (reflection spectroscopy). Thus, the main objective of this research is to develop artificial intelligence models to detect abiotic stress and estimate total chlorophyll in wheat through diffuse reflection spectroscopy. An experiment was conducted to generate fifteen different stress conditions, including lighting and water amount, from wheat seeds that germinated and grew until the seventh day. With the wheat leaf samples, reflection and absorption spectra were obtained. In the analysis for water stress classification (3ml, 5ml, and 10ml), several MiniRocket models were compared with preprocessing techniques such as the Savitzky-Golay filter, Continuum Removal (CR), De-trending (DT), Normalization of Reflectance Spectra by Area, Standard Normal Variate (SNV), and Multiplicative Scatter Correction (MSC). The results show that models using SNV and CR achieved perfect precision for validation data. For light stress classification (green, red, blue and white LEDs and one dark), models evaluated with different preprocessing techniques achieved precisions ranging from 0.841 to 0.986, where the best model obtained for light stress classification achieved a precision of 0.986 for validation data. The prediction of chlorophyll content in wheat leaves from multi-angular diffuse reflection spectroscopy was addressed using machine learning models including deep neural network (DNN) and MiniRocket, revealing that both DNN and MiniRocket achieved high performance in predicting chlorophyll content. MiniRocket showed shorter processing times (as low as 2.667 s and up to 6.840 s) but with a coefficient of determination of 0.928 for the best model. DNN had longer processing times (ranging from 76.50 s to 152.35 s) but achieved coefficients of determination of 0.993 for the validation dataset. Finally, it is concluded that the research demonstrated the effectiveness of combining multi-angular reflection spectroscopy with machine learning techniques to detect stress and estimate chlorophyll content in wheat leaves. This approach provides a more accurate and robust method, opening new perspectives and possible directions for future research in plant analysis and spectral applications.