For the long distance single photon avalanche diode (SPAD)-based underwater visible light communication (UVLC) system, the multiple-symbol union detection scheme is presented. However, an error floor curve of the bit-to-error ratio (BER) occurs and cannot vanish even though the transmitted power approaches infinity. In this paper, to solve the problems of error floor and channel estimation in the long distance SPAD-based UVLC system, we propose the one training symbol maximum likelihood (ML) detection receiver. First, we add one training symbol in the head of each frame to eliminate the error floor and ensure the reliable blind estimation of the long distance UVLC channel. Meanwhile, the training ML detection (TMLD) receiver is developed. And then, to improve the system performance, a training modified (TM) quasi-ML receiver with Anscombe root transformation is proposed. Compared to the traditional mean detection scheme of the SPAD receiver, the simulation results show that the proposed TM quasi-ML and TMLD schemes significantly improve the error-rate performance.