Microbial communities respond to physicochemical changes in the environment, making the microbiome a sensitive indicator of ecosystem status. Monitoring aquatic microbiomes is therefore essential for understanding ecosystem health and responses to change. While traditional monitoring relies on microscopy, DNA-based approaches that leverage advances in high-throughput sequencing are increasingly incorporated. Here, we evaluate the potential of using metabarcoding to predict abiotic and biotic parameters across spatiotemporal gradients of the Baltic Sea. The dataset comprises 397 seawater samples integrating prokaryotic and eukaryotic (16S and 18S rRNA gene) metabarcoding data with environmental measurements and plankton microscopy counts. Random Forest models based on metabarcoding data accurately predicted multiple physicochemical parameters and performed comparably to two other machine learning methods, XGBoost and TabPFN. Models using 16S rRNA gene data performed better than those using 18S rRNA gene data, with amplicon sequence variant-level yielding the most accurate results. Metabarcoding also exceeded plankton microscopy in predicting abiotic factors and effectively predicted the presence of phytoplankton and zooplankton genera from ≤1 L of water. Models trained on independent datasets accurately predicted several physicochemical parameters, though performance decreased for others highlighting challenges in transferability. Finally, metabarcoding-based predictions closely matched established eutrophication indicators of environmental status, demonstrating the utility of microbiome-based approaches for marine ecosystem monitoring and management.