Abstract: In order to reduce the risk of pesticide residues caused by overuse, the hyperspectral imaging technology (515-900nm) was used, and fresh tobacco leaf samples treated with three different concentrations of imidacloprid solution (1∶5000,1∶2500, 1∶850) for 24 hours were selected as the research objects. Four commonly used spectral preprocessing methods were applied, followed by inputting the preprocessed data into convolutional neural networks (CNN), random forests (RF), and least squares support vector machines (LSSVM) separately to compare their accuracies and select the optimal model across the entire spectrum. Subsequently, dimensionality reduction techniques including successive projections algorithm (SPA) and competitive adaptive reweighted sampling (CARS) were applied to the data, and the reduced hyperspectral data were inputted into the selected optimal model to establish a model for detecting imidacloprid usage. The model accuracies and the number of selected features were compared, aiming to identify the optimal feature bands. The results show that: (1) Raw data or preprocessing data input to Convolutional Neural Network, and the accuracy of test set samples result reaches 100.00%. The data preprocessed by first and second derivatives input into a random forest, and the accuracy of the testing set result also reaches 100.00%. (2) The dimensionality reduction effect of successive projections algorithm method is better than the competitive adaptive reweighted sampling method. (3) the random forest established after second derivatives preprocessing and successive projections algorithm method dimensionality reduction was ultimately chosen as the optimal discrimination model, with 12 feature wavelengths, the 98.86% of the test set result accuracy, and 0.79 ms of the single sample detection. In conclusion, the application of hyperspectral technology combined with the D2-SPA-RF model can achieve rapid detection of imidacloprid dosage in tobacco field production.