Improved chlorophyll-a estimation across large regions: a transfer learning method

Water quality is crucial to both human health and environmental sustainability. Chlorophyll-a (Chl-a), as a spectrally active compound in phytoplankton, is a key indicator of water quality. Therefore, accurate estimation of Chl-a plays an important role in water quality monitoring. Traditionally, Chl-a is measured using in situ sampling methods, which are time-consuming and labor-intensive. Recently, remote sensing, enabling high-frequency, long-term, and large-scale earth surface monitoring, has been used to estimate Chl-a based on machine learning (ML) modeling. However, ML models typically require a large number of labeled samples and have limited spatial transferability. In this study, we improve the spatial transferability of ML-based Chl-a estimation models using the strategy of transfer learning (TL). Specifically, we organized Chl-a data samples from water bodies across the United States into two domains based on the geographic locations. Meanwhile, the corresponding remote sensing observations and weather variables at each site are collected and processed as covariates. After that, datasets from two domains are interchangeably used as the training set and the testing set for the model. We prove that fine-tuning-based transfer learning (FTL) can effectively increase the transferability of a multilayer perceptron model and outperform the other baseline ML and deep learning models. An ablation study has been conducted to explore the number of layers to fine-tune and the amount of target-domain data needed to achieve the optimal performance. Finally, we analyze the importance and distribution patterns of the input features, identifying the key features and explaining the underlying mechanism of TL. We demonstrate the effectiveness of FTL in improving spatial transferability for ML-based Chl-a estimation models, and the method is scalable to other regions to address the data scarcity issue in ML-based water quality monitoring.