Assessing the ecotoxicological risk of nicosulfuron on maize using multi-source phenotype data and hyperspectral...

Herbicide-induced toxicity in maize crops poses significant challenges for agricultural management. Traditional assessment methods for herbicide toxicity in crops often show inconsistent accuracy. This study explores rapid and non-invasive techniques for evaluating herbicide toxicity, focusing on the physiological, biochemical, and growth responses of maize varieties subjected to two concentrations of nicosulfuron. We developed a comprehensive toxicity evaluation model to classify samples into three toxicity levels, showing a strong correlation (r = 0.95) with traditional tassel stage toxicity assessments. Additionally, we used hyperspectral imaging coupled with deep learning techniques to predict early toxicity levels in maize following herbicide exposure. After 4 days of herbicide treatment, our ToxicNet model using spectral data achieved an impressive 89.66 % accuracy in predicting nicosulfuron toxicity levels, facilitating early detection. Furthermore, by integrating leaf spectral data, Soil-Plant Analysis Development (SPAD) values and water content, the ToxicNet-MS model achieved a remarkable prediction accuracy of 91.38 %. Notably, this model demonstrated robust generalization across different years and planting seasons, with accuracies of 83.33 % and 89.89 %, respectively. These results significantly outperformed traditional machine learning methods (Support Vector Machine, Random Forest), classical deep learning models (Multilayer Perceptron, AlexNet), and the spectral-based ToxicNet model. This advancement offers a promising, early, and non-invasive solution for assessing herbicide-induced toxicity in maize crops, ultimately benefiting both sustainable agricultural practices and effective crop management.