The invention discloses a transform domain quadratic estimation method uniting weighted threshold de-noising and balanced decision. The method comprises the following steps: step 1), uniting a superimposed training sequence and a harmonic sequence and generating a sending sequence; step 2), in an OFDM (Orthogonal Frequency Division Multiplexing) baseband system, through an estimation method of an LS channel of the superimposed training sequence, obtaining coarse estimation of a channel frequency response; step 3), performing amplitude phase compensation on the coarse estimation of the channel frequency response, adding a window function with a width of M and performing DCT (Discrete Cosine Transformation) of an M point, performing low-pass filtering, time domain de-noising and sequence zero padding extension processing on the obtained sequence, performing N point IDCT (Inverse Discrete Cosine Transform), window removing and secondary amplitude phase compensation on an extended N point sequence, as the following formula, and obtaining the following formula; step 4), performing balanced decision on the obtained result and obtaining the following formula, repeating M point DCT, low-pass filtering, time domain de-noising, interpolation zero padding and N point IDCT, and obtaining a final estimation of a frequency domain response, as the following formula. The transform domain quadratic estimation method uniting the weighted threshold de-noising and the balanced decision provided by the invention eliminates the influence of a data sequence and the training sequence on the channel estimation performance, improves the performance of the existing LS (Least Squares) channel estimation, and has lower implementation complexity.