Effects of spectral modeling methods and land use patterns were explored on hyperspectral inversion of soil organic matter (SOM) and total phosphorus(TP) in soil samples collected from paddy fields, peach orchards and vegetable gardens in the Zhihugang catchment, Taihu Lake Region. Results show that the PLSR (Partial least square regression) model was quite high and stable in modeling and prediction precision; the GRNN (General regression neural network) of ANN (Artificial neural network) was also quite high in prediction precision, but prone to overfitting; the BPNN (Back Propagaton Neural Network) was relatively stable, but slightly low in precision; and the combined PLSR-ANN model improved in prediction precision by combining the advantages of the two in handling complicated samples. The spectral inversion of SOM was better than that of TP, and among the three patterns of land use, paddy fields were fitter than the other two for use of the models in prediction of SOM and TP. In the current study zone, patterns of land use did not have much effect on spectral inversion of SOM, but did much on that of soil TP. It is, therefore, essential to calibrate the models in light of land use patterns in conducting spectral inversion of soil TP.