Accurate and efficient acquisition of chlorophyll-a (Chl-a) concentration in water is the prerequisite for improving eutrophication and sustainable development of water bodies. This study used the hyperspectral reflectance of the water surface and the measured Chl-a water concentration as data sources. The original spectral reflectance was processed with a step size 0.1 by fractional order differentiation (FOD) technology. The characteristic bands were screened by exploring the correlation between the spectral data and the Chl-a concentration of the water body, and the variable-order spectral dataset was constructed. The Partial Least Squares (PLS) model was used to screen the optimal features to construct the dataset, which was divided into modeling set and verification set according to the ratio of 7∶3, and the support vector machines (SVM) and deep forest (DF) models were used to establish the water Chl-a concentration inversion model. It is also compared with the model constructed using the original data and the model constructed by common dimensionality reduction methods. The results show that FOD technology can reduce hyperspectral noise, mine potential spectral information to a certain extent, and improve the correlation between hyperspectral reflectance and the concentration of Chl-a in water. Compared with the Chl-a inversion model established by using the original data and PCA dimension reduction, the R2 of the water Chl-a concentration inversion model established by FOD combined with PLS first screening features was increased, and the mean square error (MSE) and mean absolute error (MAE) were reduced. DF has a higher degree of fitting and prediction accuracy than the other three models, with R2=0.96, MSE=0.51 μg·L-1, and MAE=0.64 μg·L-1. The validation set R2=0.89, MSE=0.69 μg·L-1, MAE=0.64 μg·L-1. In general, it is feasible to establish a water Chl-a concentration inversion model based on the variable-order spectral dataset after FOD recombination and the preferred features of PLS. Comparative analysis of the inversion results of other models shows that DF has a good inversion effect on water Chl-a. This work provides a theoretical basis and technical support for the inversion of hyperspectral remote sensing of Chl-a in inland second-class water bodies, helps the continuous monitoring of water quality in Baiyangdian Lake, and also provides new ideas for the inversion of Chl-a from hyperspectral satellite remote sensing images in the future.