The effect of Mixed-hyperspectral in the water is difficult in quantitative remote sensing of water. Studies have shown that the only scalar spectrum information is difficult to solve the problem of complex mixed spectra of water. Besides the spectral information, spatial distribution of information is one of the obvious characteristics of the broad waters pollution, and can be used as a useful complement to the remote sensing information and facilitate water complex spectral unmixing. Taking Chaohu as an example, the paper applies the HJ-1A HSI hyperspectral data and the supplemental surface spectral measurement data to discuss the mixed spectra of lake water by spatial statistics and genetic algorithm theory. By using the spatial variogram of geostatistics to simulate the distribution difference of two adjacent pixels, the space-informational decomposition model of mixed spectral in lake water is established by co-kriging genetic algorithm, which is a improved algorithm applying the spatial variogram function of neighborhood pixel as the constraint of the objective function of the genetic algorithm. Finally, the model inversion results of suspended matter concentration are verified. Compared with the conventional spectral unmixing model, the results show the correlation coefficient of the predicted and measured value of suspended sediment concentration is 0.82, the root mean square error 9.25 mg·L-1 by mixed spectral space information decomposition model, so the correlation coefficient is increased by 8.9%, the root mean square error reduced by 2.78 mg·L-1, indicating that the model of suspended matter concentration has a strong predictive ability. Therefore, the effective combination of spatial and spectral information of water, can avoid inversion result distortion due to weak spectral signal of water color parameters, and large amount of calculation of information extraction because of the high spectral band numbers, and also provides an effective way to solve spectral mixture model of complex water and improve the accuracy of model inversion.