The response characteristics of TM image to the soil moisture in the Tarim River are the research object. Selected the image spectrum (R), spectrum reciprocal (1/R), the logarithm of reciprocal spectrum lg(1/R) and removal normalized difference vegetation index (R-c) of four spectral index were selected to establish the soil moisture content prediction model, the variance test was used to validate the model significance, the model accuracy level was divided by the posterior variance examination. The results showed that: the model accuracy of the logarithm of reciprocal spectrum lg(1/R) prediction of soil moisture is the highest, and achieved a good level for the monitoring of soil moisture content (0 similar to 30 cm). The model accuracy of the spectral (R) and spectral reciprocal (1/R) prediction of soil moisture is lower than logarithm of reciprocal spectrum with only the individual layers (0 similar to 30, 0 similar to 50 cm, etc.) reaching the qualified level or narrouly qualified level. The model accuracy of the removal normalized difference vegetation index (R-c) prediction of soil moisture is the lowest. Besides, the best prediction depth of every model is the depth of 0 similar to 30 cm, and if the soil depth is too deep or too shallow, the prediction accuracy will decrease.