Online detection of fertilizer information based on Raman spectroscopy and deep learning and the design of detection system

Abstract Online detection of fertilizer information is crucial for achieving intelligent and precise fertilizer application. However, traditional methods face challenges such as difficulty in detecting specific concentrations of ions and cross-contamination. In this study, Raman spectroscopy was combined with deep learning, to identify nine types of fertilizers: K2SO4, KNO3, KH2PO4, (NH4)2HPO4, CO(NH2)2, CO(NH2)2 + KH2PO4, K2SO4 + KH2PO4, K2SO4 + KNO3, and (NH4)2HPO4 + KNO3, along with their ion concentrations. Qualitative and quantitative models of fertilizers were constructed using three deep learning algorithms combined with a feature band selection method, a feature wavenumber extraction method, and eight spectral preprocessing methods. The gated recurrent unit (GRU) under Savitzkt-Golay (SG) smoothing, baseline correction (BC), and de-trending (DT) preprocessing had the best classification performance, with its precision, recall, and F1 scores reaching more than 98.29% in the test set. Moreover, in terms of regression, GRU outperformed one-dimensional convolutional neural network (1D-CNN) and long short-term memory network - transformer (LSTM-Transformer), yielding prediction R2 within the range of 0.9716–0.9977 and a root mean square error in the range of 0.1472–0.4713 in the test set. Building on this foundation, software and hardware were designed to construct an online detection system for fertilizer solutions. The test results showed that the system had an average accuracy of 97.78% in fertilizer types identification, and the error ranges of ion concentration prediction were all within ±7.95%. This study addresses the practical requirements for online fertilizer detection in agricultural engineering, laying the foundation for the realization of water-fertilizer integration.