Marine ecological environment monitoring is shifting from single-index empirical judgment to dynamic identification and trend prediction driven by multi-source observation data. In order to solve the problems of strong coupling, significant temporal fluctuation, and prominent spatial heterogeneity of ecological variables in typical inshore waters, and it is difficult for traditional methods to simultaneously consider state discrimination and continuous prediction, this paper takes the inshore waters of Jiaxing and Ningbo as objects, and constructs a machine learning model combining sliding time window, spatial neighborhood weight, unified expression of multi-source features and joint output of dual tasks. The model takes remote sensing observation, bubue monitoring, water quality investigation, meteorological and Marine data and human activity information as input, and realizes a comprehensive description of the evolution process of Marine ecology through spatio-temporal feature fusion, state semantic mapping, category discrimination and trend prediction. The experimental results show that the contribution of inorganic nitrogen in Jiaxing sea area reaches 0.24, and the proportion of degraded and high-risk samples is 52.9%. The RMSE of the model under 1-step and 12-step prediction conditions is 0.028 and 0.086, respectively, the F1 value is still 0.833 under 40% missing rate, and the accuracy of cross-region transfer reaches 0.887 and 0.873, which is better than that of XGBoost, LSTM and SVR. The research can provide effective technical support for typical sea ecological environment monitoring, risk early warning and fine governance.
