Deep-learning-driven spectral image analysis for intelligent monitoring of multiple pesticides and antibiotics.

Abstract

With the widespread use of pesticides and antibiotics in agriculture and healthcare, their associated environmental pollution and potential health hazards have emerged as a global concern. This study presents a novel deep learning-based spectral image analysis approach that is dedicated to the intelligent monitoring of multiple pesticides and antibiotics in agricultural water bodies. A total of 6100 samples containing glyphosate (GL), bentazone (BE), benzylpenicillin potassium (BP), and tetracycline hydrochloride (TH) at concentrations range of 3.8-550 μg/L were prepared. After the samples were mixed with selected composite chromogenic reagents, the specific absorbance characteristics of the stabilized reaction mixtures were measured using a custom-designed spectrometer. The preprocessed spectral data were used to train a fine-tuned ResNet-50 deep learning model. By establishing mappings between spectral features and reference concentrations, the model effectively predicted unknown pollutant concentrations. The results indicated that the proposed method enables rapid and simultaneous detection of GL, BE, BP and TH. Under laboratory conditions, the coefficient of determination exceeded 0.993, the reliable prediction rate was over 80 % in the concentration range of 10-550 μg/L. The limits of detection for GL, BE, BP, and TH were 0.23, 0.32, 0.38, and 0.28 μg/L, respectively. In addition, the frequency of abnormal predictions for natural water samples exhibited an increase over the concentration range of 3.8-10 μg/L, while the overall accuracy remained relatively high. Our research provides a new perspective on the rapid identification of pesticides and antibiotics. In the future, we hope this method can offer a timely, cost-effective and scalable solution for the early warning and real-time tracking of pollutants in water bodies.