Efficient and Non-Invasive Evaluation of Nicosulfuron Toxicity Using Hyperspectral Imaging and Machine Learning

Abstract

Toxicity from herbicides on crops has become a widespread global problem. Rapid and non-invasive methods for evaluating herbicide toxicity have not been fully explored. This study evaluated the toxicity of 15 maize varieties under two concentrations of nicosulfuron using hyperspectral imaging and machine learning. Typical morphological, physiological, and biochemical indicators were measured 14 days after herbicide application to analyze the toxicity mechanism of nicosulfuron. A comprehensive toxicity evaluation model was developed to evaluate toxicity and classified 30 samples into three levels of toxicity. Spectral analysis showed that the spectral differences were most significant on the first day after nicosulfuron application, and the toxicity was mitigation over time. Differences were observed within the spectral ranges of 495-679 nm and 652-716 nm. Support vector machsine (SVM), random forest (RF), and partial least squares discriminant analysis (PLS-DA) were used for toxicity prediction. The highest accuracy achieved 70.65% for all toxicity levels and 81.25% for mild and severe toxicity (grade I and III) was achieved on days 1 and 4 post-niclosulfuron application. Our results indicated that Visible-NIR HSI technology can be used to analyze nicosulfuron toxicity status and predict early toxicity levels, providing an efficient and non-invasive method to evaluate herbicide toxicity in crops.

Keywords: Herbicide toxicity, Hyperspectral imaging, Machine learning, Maize, Nicosulfuron