5:15 PM - 7:15 PM
[ACG50-P04] Estimation of global distribution of seawater dimethyl sulfide (DMS) concentration using machine learning
Keywords:seawater DMS concentration, artificial neural network, phytoplankton
1. Introduction
Since emission of dimethyl sulfide (DMS) from the sea surface to atmosphere is one of key processes to affect the microphysical and chemical properties of aerosols, it is required to be accurately estimate the spatial and temporal distributions of seawater DMS concentration and its emission to the atmosphere. Thus far, there have been several approaches to estimate the global distribution of seawater DMS based on the regression analysis (Kettle et al., 1999; Lana et al., 2011; Gali et al., 2018). However, because of the complexities of production and loss processes of seawater DMS, uncertainties of these estimation are still large (Tesdal et al., 2016). Recently, there are several reports on the estimation of seawater DMS using machine learning such as Artificial Neural Network (ANN) (Mansour et al., 2023; McNabb and Tortell, 2022; Wang et al., 2020). It has been reported that the accuracy of estimation of DMS concentrations using these machine learning is considerably improved as compared to the results using regression models. However, in these methods, information of phytoplankton species, which are known to be a significant effect on DMSP production, are not explicitly considered as predictor variables. Therefore, in order to improve the accuracy, I am trying to estimate the distribution of seawater DMS concentrations by incorporating phytoplankton species information into the predictor variables of ANN, and to investigate the variations in seawater DMS concentration distribution and its emission from 2003 to 2022.
2. Dataset and Methods
ANN was used to estimate the daily seawater DMS concentration from 2003 to 2022 with the resolution of 0.25 degree in longitude and latitude. For the training and validation for the ANN, in-situ seawater DMS concentrations derived from the Global Surface Seawater DMS Database (Kettle et al., 1999) and the NAAMES (Behrenfeld et al., 2019) were used. For predictor variables of ANN, following dataset were used; ocean color data derived from MODIS-Aqua, mixed layer depth from MIMOC and nutrients data such as nitrate and phosphate from Global Ocean Biogeochemistry Hindcast of Copernicus Marine Service. The phytoplankton species were analyzed using the PHYSAT algorithm (Alvain et al., 2005; 2008; 2012) to obtain the dominant species among 6 species and the algorithm developed by Hirata et al. (2011) to obtain the Chl-a concentration of 8 phytoplankton functional types. Base on the accuracy assessment using RMSE, the best results so far have been achieved with 9 predictor variables, 2 hidden layers, and 26 and 13 nodes in each hidden layer, respectively.
3. Results and summary
The 20-year monthly mean distribution of DMS concentrations calculated from ANN was compared with those of climatological distribution of DMS Rev3 (Hulswar et al., 2022) (Figure). The similarity between the two is that high concentrations of DMS are calculated in the high latitudes in the summer months in the both hemispheres, while the difference between them is that high concentrations are calculated in the Indian Ocean in the summer months in DMS Rev3, which is not calculated in my ANN result. Global DMS emission into the atmosphere was calculated using the wind speed at the height of 10m above sea surface and SST. Mean global DMS emission for this period was estimated to be 22.6 TgS yr-1, and about 17% less than that (27.1 TgS yr-1) calculated from the climatology data of DMS Rev3. There is still room for improvement in terms of accuracy for the current training data, and I am trying to further improve the accuracy by changing the combination of predictor variables and the number of hidden layers and nodes. In addition, I will try to investigate the relationship between DMS emissions and aerosol concentrations using proxy data such as aerosol index over selected ocean areas.
Since emission of dimethyl sulfide (DMS) from the sea surface to atmosphere is one of key processes to affect the microphysical and chemical properties of aerosols, it is required to be accurately estimate the spatial and temporal distributions of seawater DMS concentration and its emission to the atmosphere. Thus far, there have been several approaches to estimate the global distribution of seawater DMS based on the regression analysis (Kettle et al., 1999; Lana et al., 2011; Gali et al., 2018). However, because of the complexities of production and loss processes of seawater DMS, uncertainties of these estimation are still large (Tesdal et al., 2016). Recently, there are several reports on the estimation of seawater DMS using machine learning such as Artificial Neural Network (ANN) (Mansour et al., 2023; McNabb and Tortell, 2022; Wang et al., 2020). It has been reported that the accuracy of estimation of DMS concentrations using these machine learning is considerably improved as compared to the results using regression models. However, in these methods, information of phytoplankton species, which are known to be a significant effect on DMSP production, are not explicitly considered as predictor variables. Therefore, in order to improve the accuracy, I am trying to estimate the distribution of seawater DMS concentrations by incorporating phytoplankton species information into the predictor variables of ANN, and to investigate the variations in seawater DMS concentration distribution and its emission from 2003 to 2022.
2. Dataset and Methods
ANN was used to estimate the daily seawater DMS concentration from 2003 to 2022 with the resolution of 0.25 degree in longitude and latitude. For the training and validation for the ANN, in-situ seawater DMS concentrations derived from the Global Surface Seawater DMS Database (Kettle et al., 1999) and the NAAMES (Behrenfeld et al., 2019) were used. For predictor variables of ANN, following dataset were used; ocean color data derived from MODIS-Aqua, mixed layer depth from MIMOC and nutrients data such as nitrate and phosphate from Global Ocean Biogeochemistry Hindcast of Copernicus Marine Service. The phytoplankton species were analyzed using the PHYSAT algorithm (Alvain et al., 2005; 2008; 2012) to obtain the dominant species among 6 species and the algorithm developed by Hirata et al. (2011) to obtain the Chl-a concentration of 8 phytoplankton functional types. Base on the accuracy assessment using RMSE, the best results so far have been achieved with 9 predictor variables, 2 hidden layers, and 26 and 13 nodes in each hidden layer, respectively.
3. Results and summary
The 20-year monthly mean distribution of DMS concentrations calculated from ANN was compared with those of climatological distribution of DMS Rev3 (Hulswar et al., 2022) (Figure). The similarity between the two is that high concentrations of DMS are calculated in the high latitudes in the summer months in the both hemispheres, while the difference between them is that high concentrations are calculated in the Indian Ocean in the summer months in DMS Rev3, which is not calculated in my ANN result. Global DMS emission into the atmosphere was calculated using the wind speed at the height of 10m above sea surface and SST. Mean global DMS emission for this period was estimated to be 22.6 TgS yr-1, and about 17% less than that (27.1 TgS yr-1) calculated from the climatology data of DMS Rev3. There is still room for improvement in terms of accuracy for the current training data, and I am trying to further improve the accuracy by changing the combination of predictor variables and the number of hidden layers and nodes. In addition, I will try to investigate the relationship between DMS emissions and aerosol concentrations using proxy data such as aerosol index over selected ocean areas.