Integer-order derivative methods (such as 1st or 2nd order) are traditional preprocessing methods for soil heavy-metal inversion models, which ignore the fractional-order spectral reflectance information associated with the target variable. Fractional order derivative (FOD) can flexibly select the differential order to enhance the spectral signal effectively. This study focused on the farmland soil in Mojiang Hani Autonomous County, Pu’er City, Yunnan Province, China. Sixty-one soil hyperspectral reflectance information and soil heavy metal content data (zinc and nickel) were measured. The spectral reflectance information underwent 0 to 2 fractional-order derivative preprocessing with intervals of 0.05. The preprocessed spectral reflectance information at each order was input into the Successive Projections Algorithm (SPA) to select characteristic bands. Subsequently, three soil heavy metal prediction models were separately established using Partial Least Squares Regression (PLSR), Random Forest (RF), and Bagging methods. The results show that after the fractional order derivative processing from 0 to 2 orders (41 orders in total with an interval of 0.05), the overall spectral intensity gradually weakens and gradually approaches zero with the increase of fractional orders. The spectral absorption band gradually narrows, and the differences between different spectral curves gradually decrease. As the derivative order increases, more abundant peaks and valleys are produced. The best-order models based on fractional derivatives are better than the original spectral model and the integer order model, and most of the better orders of the model are concentrated in low-order fractional orders. For heavy metal zinc, the best prediction model accuracy was achieved by the RF model of 0.75 order (R²=0.675, RMSE=6.149, RPD=1.755), followed by the Bagging model of 0.75 order (R²=0.633, RMSE=6.534, RPD=1.652), and the lowest was achieved by the PLSR model of 0.25 order (R²=0.551, RMSE=7.230, RPD=1.493). For the heavy metal nickel, the best prediction model accuracy was the RF model of order 0.80 (R²=0.854, RMSE=127.823, RPD=2.618), the Bagging model of order 0.80 was the next best (R²=0.841, RMSE=133.304, RPD=2.510), the PLSR model of order 0.40 lowest (R²=0.762, RMSE=163.162, RPD=2.051). Visible, the nonlinear models (RF and Bagging) constructed based on FOD preprocessing and SPA dimensionality reduction in this study have certain applicability in estimating heavy metal content in farmland soil. They can be a reference for predicting heavy metal content in similar regions.