Coastal ecosystems are evolving with the increase of anthropogenic activities. Their dynamics involve various spatial and temporal scales, as well as complex benthic and pelagic interactions. Understanding these dynamics necessitates further knowledge of marine extreme, recurrent, and rare events, e.g., heat waves, Harmful Algal Blooms (HABs), storms, flood, etc.  Thus, the development of a forecasting system that alerts for algal blooms and other environmental states becomes imperative inorder to  mitigate their socio-economic and environmental influences.  In this research, we developed a semi-supervised machine learning approach to forecast marine environmental states, including algal blooms. Our approach is a multi-source, multi-frequency, and multi-parameter approach that involves in-situ, satellite and modeling data,  at low and high frequency. We apply the unsupervised M-SC (Multi-level Spectral Clustering) algorithm to cluster the data both spatially and temporally. Following that, we label these clusters to characterize the different environmental states, such as rare, extreme and recurrent events. Then, we apply a supervised machine learning algorithm such as Random Forest (RF) in order to forecast future environmental states, particularly algal blooms. This expert system will lead to better management strategies for marine ecosystems, and will help mitigate algal blooms.