Vertical Chlorophyll-a Profile Reconstruction in the North Indian Ocean Using Deep Neural Networks

To gain an in-depth understanding of the distribution patterns of marine primary productivity and support the monitoring of marine ecological environment changes, this paper integrates Biogeochemical-Argo (BGC-Argo) float data with satellite remote sensing data to construct a deep learning model for retrieving the vertical distribution characteristics of Chlorophyll-a (Chl-a) mass concentration in the North Indian Ocean. The model combines the vertical temperature and salinity profile features extracted by the 1DCNN-Transformer module, the spatiotemporal continuous vectors transformed by the Embedding module, and satellite remote sensing data, which are then fed into a Deep Neural Network (DNN) to reconstruct the vertical distribution of Chl-a mass concentration in the North Indian Ocean. Sensitivity analysis identified Sea Surface Temperature (SST), sea surface Chl-a mass concentration, and vertical temperature and salinity profiles as the optimal input variables for the model. The results show that the model achieves an RMSE of 0.106 μg/L and an R² of 0.81 in reconstructing Chl-a mass concentration profiles. Within the 0-140 m depth range, the maximum RMSE for Chl-a concentration retrieval is approximately 0.4 μg/L, while errors decrease significantly below 140 m. In terms of spatial distribution, the maximum Chl-a concentration (A) in the Arabian Sea is significantly higher than that in the Bay of Bengal and equatorial regions. In terms of temporal variation, the A values and their corresponding depths (Z_max) in the Arabian Sea, Bay of Bengal, and equatorial regions exhibit seasonal variations, with a certain negative correlation observed between Z_max and A.