For decades, the global remotely sensed significant wave heights have been from altimeters and/or synthetic aperture radars in wave mode, which both suffer from spatial and temporal sampling limitations. In contrast, spaceborne scatterometers are with large swath and high temporal revisit frequency at a global scale, but so far are routinely providing ocean winds rather than waves. This paper addresses the ocean sea state retrieval algorithm by applying state-of-the-art machine learning technology to European Advanced Scatterometer (ASCAT). A huge collocation database (< 6 million) has been built between L1b/L2 ASCAT products and WaveWatch III (ww3) ocean wave hindcasts within the spatio-temporal criteria of 0.1 degree and 0.5 h for the period of three years, followed by the mining of this big data by means of machine learning (i.e., multi-hidden layer neural network here). The neural network proposed here includes layers: the input layer (13 ASCAT variables), four hidden layers, and the output layer (wave heights). The performance of machine learning based approach for ocean wave height estimation from scatterometer is evaluated using two independent match-ups: ASCAT-WW3 and ASCAT-buoy. The statistical assessment against SWH hindcast shows the root mean square error of 0.55 m and scatter index of 23%, respectively. Results indicate that the data driven algorithm is reasonable for sea state estimation from wide-swath scatterometers, and encouraging for operational implementation in the future.