In quantitative analysis of spectral data, noises and background interference always degrade the accuracy of spectral feature extraction. The wavelet transform is multi-scale decomposition used to reduce the noise and improve the analysis precision. On the other hand, the wavelet transform denoising is often followed by destroying the efficiency information. The present research introduced two indexes to control the scale of decomposition, the smoothness index (SI) and the time shift index (TSI). When the parameters satisfied TSI <0.01 and SI>0.100 4, the noise of spectral characteristic was reduced. In the meanwhile, the reflection peaks of biochemical components were reserved. Through analyzing the correlation between denoised spectrum and chlorophyll content, some spectral characteristics parameters reflecting the changing tendency of chlorophyll content were chosen. Finally, the partial least squares regression (PLSR) was used to develop the prediction model of the chlorophyll content of tomato leaf. The result showed that the predictiong model, which used the values of absorbance at 366, 405, 436, 554, 675 and 693 nm as input variables, had higher predictive ability (calibration coefficient was 0.892 6, and validation coefficient was 0.829 7) and better potential to diagnose tomato growth in greenhouse.