A chlorophyll-a spectrum index (CSI) is built based on the linear spectral unmixing method using the endmembers selected from in situ spectra. The character of CSI is analyzed based on a dataset including 307 samples collected from the Taihu Lake, Chaohu Lake, Dianchi Lake, and the Three Georges Reservoir. Furthermore, a chlorophyll-a concentration (Cchla) estimation model is built and compared with traditional models in terms of noise immunity and sensor adaptability. The results show that: 1) CSI is a good indicator for Cchla. Two spectrum sets that are divided by f_CSI=0 clearly show different spectrum characteristics; 2) for the hyperspectral dataset, the CSI algorithm gets similar performance to that of the three band algorithm (TBA) (their validation mean absolute percent errors are 0.332 and 0.330, respectively, root mean squares errors are 9.892 and 9.929, respectively); 3) the CSI algorithm is not sensitive to both unbiased and biased noise. Its accuracy is nearly independent on the unbiased noise. Meanwhile, the traditional three band algorithm is sensitive to both of the two kinds of noise; 4) the CSI algorithm is less sensitive to the band settings of the remote sensor than the traditional algorithms. The advantage is more obvious for wide band multispectral sensors. Compared with traditional Cchla semi-empirical models, the CSI algorithm is more stable and potential.