Vegetable heavy metal pollution monitoring is an essential component of precision agriculture. In order to explore the hyperspectral response of vegetable under different concentrations of heavy metal stress, hyperspectral imaging (HSI) data of cabbage leaves under ten concentrations of Cu2+ stress was collected, and feature wavelength selection and classification modeling was conducted. A convolutional long shortterm memory neural network (CNN-LSTM) model for cabbage leaf Cu2+ stress classification based on hyperspectral data was proposed. Ten cabbage samples with different concentration gradients of copper stress were set, with four pots of samples at each concentration. Hyperspectral data collection was carried out when the cabbage grew to 15~20 leaves. Ten leaf samples were collected for each cabbage. A total of 400 hyperspectral data were collected. Firstly, spectral data preprocessing was performed by using S-G smoothing and first-order differentiation. Then, competitive adaptive reweighted sampling (CARS) and uninformative variable elimination (UVE) were used to extract ten common feature bands. The experimental results indicated that using the common wavelengths extracted by the UVE and CARS methods as input for the CNN-LSTM model achieved a test set accuracy of 94.8%, precision of 93.1%, and recall of 93.5%. These values were higher than those achieved by the SVM, CNN, and LSTM models by 8.7, 5.7, and 6.4 percentage points in accuracy, 6.6, 4.7, and 5.9 percentage points in precision, and 10.1, 5.2, and 3.9 percentage points in recall, respectively. The results were verified by accurate measurement of heavy metal content in cabbage leaves using an ICP-700T inductively coupled plasma emission spectrometer. For non-destructive classification monitoring of cabbage leaves under copper stress, the CNN-LSTM classification model with UVE-CARS feature bands selection performed the best, providing a method for non-destructive detection of vegetable heavy metals.