B1D–CNN: A Novel Convolution Neural Network-Based Chlorophyll-A Retrieval Algorithm for Sentinel–2 Data

Deep learning (DL) methods have been recently considered suitable for Chlorophyll-a (Chla) retrieval from satellite data due to their ability for handling complex, high-dimensional, and noisy data. This manuscript describes B1D–CNN, a new model combining 1D-convolutional neural networks (1D-CNN) and a traditional empirical blend algorithm to estimate Chla using the MultiSpectral Instrument (MSI) sensor on board the Sentinel-2 satellite. The proposed model is trained and evaluated against the state-of-the-art Mixture Density Network (MDN) and the classical Chla retrieval algorithms using global in-situ data. The results show a 9.25% to 54.12% improvement in the RMSE, along with a 3.45% to 59.53% reduction in the MAE. A satellite image during a harmful algal bloom (HABs) event was also assessed using B1D–CNN, and the proposed model captured the high Chla batches associated with HABs. This study indicates the advantages of using DL methods to retrieve Chla.