3LATNet: Attention based deep learning model for global Chlorophyll-a retrieval from GCOM-C satellite

Abstract

Chlorophyll-a (Chla) retrieval from satellite observations is crucial for assessing water quality and the health of aquatic ecosystems. Utilizing satellite data, while invaluable, poses challenges including inherent satellite biases, the necessity for precise atmospheric correction (AC), and the complexity of water bodies, all of which complicate establishing a reliable relationship between remote sensing reflectance (Rrs) and Chla concentrations. Furthermore, the Global Change Observation Mission − Climate (GCOM-C) satellite operated by Japan Aerospace Exploration Agency (JAXA) has brought a significant leap forward in ocean color monitoring, featuring a 250 m spatial resolution and integrating the 380 nm band, enhancing the detection capabilities for aquatic environments. JAXA’s standard Chla product grounded in empirical algorithms, coupled with the limited research on the impact of atmospheric correction (AC) on Rrs products, underscores the need for further analysis of these factors. This study introduces the three bidirectional Long short–term memory and ATtention mechanism Network (3LATNet) model that was trained on a large dataset incorporating 5610 in-situ Rrs measurements and their corresponding Chla concentrations collected from global locations to cover broad trophic status. The Rrs spectra have been resampled to the Second-Generation Global Imager (SGLI) aboard GCOM-C. The model was also trained using satellite matchup data, aiming to achieve a generalized deep-learning model. 3LATNet was evaluated compared to conventional Chla algorithms and ML algorithms, including JAXA’s standard Chla product. Our findings reveal a remarkable reduction in Chla estimation error, marked by a 42.5 % (from 17 to 9.77 mg/m3) reduction in mean absolute error (MAE) and a 57.3 % (from 43.12 to 18.43 mg/m3) reduction in root mean square error (RMSE) compared to JAXA’s standard Chla algorithm using in-situ data, and nearly a twofold improvement in absolute errors when evaluating using matchup SGLI Rrs. Furthermore, we conduct an in-depth assessment of the impact of AC on the models’ performance. SeaDAS predominantly exhibited invalid reflectance values at the 412 nm band, while OC-SMART displayed more significant variability in percentage errors. In comparison, JAXA’s AC proved more precise in retrieving Rrs. We comprehensively evaluated the spatial consistency of Chla models under clear and harmful algal bloom events. 3LATNet effectively captured Chla patterns across various ranges. Conversely, the RF algorithm frequently overestimates Chla concentrations in the low to mid-range. JAXA’s Chla algorithm, on the other hand, consistently tends to underestimate Chla concentrations, a trend that is particularly pronounced in high-range Chla areas and during harmful algal bloom events. These outcomes underscore the potential of our innovative approach for enhancing global-scale water quality monitoring.