Monitoring Coastal Water Quality and Marine Ecosystem Change Using Satellite Remote Sensing and Machine Learning

Coastal marine ecosystems are critical for fisheries, carbon sequestration, and the livelihoods of billions. They face increasing pressure from eutrophication, ocean warming, and increased turbidity, all of which are intensifying the frequency and severity of harmful algal blooms (HABs). Traditional in-situ water sampling is too slow, costly, and spatially sparse to track these dynamics at scale, particularly for remote or low-income coastlines. This tutorial demonstrates how satellite remote sensing and machine learning can close that gap, using three years of daily chlorophyll-a and sea surface temperature (SST) observations from the Copernicus Marine Service over the Northern Adriatic Sea (a region with well-documented increases in HAB activity). Starting from exploratory data analysis, the tutorial builds classical ML models (Random Forest, Gradient Boosting) that predict optical bloom indicators (Chl-a, NDCI, and a turbidity proxy) from physical drivers alone (SST, seasonality, and location) before progressing to a patch-based convolutional neural network that incorporates spatial SST context. On a held-out 2022 test year, models achieve R² values of 0.48-0.65 (Chl-a), 0.39-0.62 (NDCI), and 0.44-0.70 (turbidity), with the CNN outperforming pixel-wise models on Chl-a (0.65 vs 0.48) by exploiting spatial thermal gradients near the Po River plume. The results show that physical drivers alone carry predictive signal for bloom preconditions - useful on cloud-covered days when direct ocean-colour observations are unavailable. The tutorial also surfaces key limitations (dominance of static spatial predictors, underprediction of extreme bloom events) that motivate future work incorporating river discharge, wind stress, and longer training records.