Abstract:The paper makes an attempt to take ’the mean square of inversion errors=minimum’ as a rule to find out an optimum subset of channels with a few different regression methods in order to set up an optimum least square inversion equation for atmospheric remote sensing,thus avoiding the generation of unstable regression coefficients due to ill-conditioned computations. The effects of regression equations resulting from such regression methods as the optimum subset method,forward method,backward method,stepwise regression (double-F-tested), single-F-tested regression,and fixed-M method,are compared with each other based on actual examples. It is shown that the last two methods suggested in the paper are not only simple in computation and distinct in meaning,but also give satisfying results.