Iron oxide is the main body of iron-bearing minerals in the soil and is one of the most obvious and useful indicators of soil development and soil classification. In this paper, forest soils in Dawei Mountain of Hunan Province were collected, iron oxide contents in soils were determined respectively by conventional chemical method and by hyperspectral inversion with models of multiple stepwise regression and partial least squares regression inversion established by screening sensitive bands after spectral preprocessing and combinatorial transformation. The results showed that soil spectral curves with different iron oxide contents all were in steep-hill shape in the whole band, iron oxide content was negatively correlated with spectral reflectance within 420-580 nm, different spectral transformation could improve the correlation, and the combination of Savitzky-Golay (S-G) smoothing and de-embedding lines was superior to other pretreatment methods in inversion. The characteristic bands of iron oxides were 392, 427, 529, 523, 549, 559, 565, 570, 994 and 1 040 nm. Partial least squares regression model had better stability than multiple stepwise regression model, and is suitable for rapid estimation of iron oxide contents in forest red and yellow soils.