Phytoplankton represents an important indicator of the ecological state of the marine environment. As in-situ measurements are difficult to obtain and time-consuming to analyze, we explored alternative methods to estimate phytoplankton pigment concentrations from other environmental parameters and investigated the feasibility of estimating phytoplankton pigment concentrations in the Gulf of Trieste using machine learning models. A set of ecologically relevant features was selected in collaboration with domain experts to construct the input dataset, which was preprocessed including appropriate handling of missing values. The study employs basic regression models and neural networks, and since real pigment measurements are scarce, we employed a transfer learning strategy: we pre-trained a model on satellite-derived chlorophyll-a data and fine-tuned it with ground measurements. For this purpose, multilayer perceptrons (MLPs) and convolutional neural networks (CNNs) were applied. The application of transfer learning led to an improvement in predictive performance, demonstrating its potential to enhance phytoplankton pigment estimation in data-limited marine environments.
|
