In the “natural farmland ecosystem”, although the biochemical parameters of crops are abnormal under the stress of heavy metal pollution, their characteristics are often very weak, with small changes and extreme unstability. Wavelet analysis, a common signal processing method in unstable signal processing, was used to process spectral reflectance data of crops (rice) and effectively extract weak information of “mutation” hidden in spectral signals under the stress of heavy metal pollution. Wavelet transform was carried out by using Db-5 wavelet basis, and singular points with abnormal spectral characteristics were selected. Back propagation neural network model (BPNN) was constructed by using spectral reflectance of corresponding bands of singular points (716nm, 745nm and 766nm) to invert the contents of four heavy metals in rice canopy. Correlation analysis was conducted between the predicted and measured values of the model, and the results showed that the inversion model of heavy metal content in rice canopy based on BP neural network had a good inversion effect on the stress of cadmium, lead, mercury and arsenic in the experimental area.