Endmember extraction is the key procedure for spectral unmixing of hyperspectral remote sensing image. In the linear spectral mixture analysis, a convex invariance of simplex was introduced when hyperspectral remote sensing image was projected into null space of spectral signature matrix of endmembers. On the basis of the invariance, a null space spectral projection algorithm (NSSPA) was proposed. One-unit endmember extraction strategies were established to implement the algorithm in a flexible way by designing different metrics and principles. It is proved that the proposed algorithm extends the algorithm based on subspace projection distance, including the classical orthogonal subspace projection (OSP) algorithm and the null space maximal distance algorithm. The algorithm provides diversified strategies for endmember extraction. The experimental results indicate that NSSPA demonstrates excellent performance of endmember extraction both in the simulated and real hyperspectral remote sensing images.