Global Chlorophyll-? Concentration Estimation From VIIRS Using Deep Learning Methods

Chlorophyll-a(chla)is the primary pigment of phytoplankton and the indicator of marine biomass.Satellite VIIRS will provide global ocean color data for the next 20 years and it is of great significance for scientific research to estimate global chla concentration using remote sensing technology timely and accurately.Deep learning has great potential in estimating global chla concentration accurately and timely for its strong modeling and computing capabilities,however,no studies focus on using deep learning methods to estimate global chla concentration.In this study,deep learning methods are employed to estimate global chla concentration from VIIRS and we implement spatial and temporal analysis of global chla concentration.The main work and conclusions are as follows:(1)In order to estimate chla concentration accurately and efficiently,a chla estimation model based on CNN is built,which the inputs are patches consist of four ocean color reflectance bands and OC-CCI chla concentration image is used as ground truth to build an end-to-end model.After fine adjustment of parameters,twelve monthly global chla concentration maps in 2017 are generated by the trained CNN model.Qualitative and quantitative analysis is used to evaluate the results and the shallow model SVR is compared with the CNN model.Comparing with the SVR,the CNN performs better with the mean log R~2 and RMSE of being 0.96 and 0.11respectively.These results demonstrate that the CNN model may provide chla concentration images reliably,timely and stably.(2)In order to solve the problem of missing data in the results of CNN model,a chla missing data interpolation model based on RNN is built,which using the published OC-CCI data.Through comparative analysis of common structures and fine adjustment of parameters,the GRU structure is determined to be the final model.The input time series length of the model is19 years which is the monthly chla concentration from 1998 to 2016,and the the predicted time series length is one year which is the monthly chla concentration of 2017.The R~2 and RMSE are 0.94 and 0.07 on the test dataset.These results demonstrate that the RNN model can interpolate precisely most of the missing data estimated by the CNN model.(3)In order to understand the dynamics of phytoplankton growth and global carbon cycle,spatial and temporal analysis of global chla concentration needs to implement efficiently and accurately.Monthly chla concentration of 2017 generated by the CNN model and RNN model are used.Spatial and temporal analysis is carried out from four aspects:annual average analysis,monthly average analysis,seasonal influence of latitude gradient and concentration change in different research areas.The results show that the distribution of chlorophyll a concentration is related to offshore distance,ocean depth and ocean current,and the maximum chlorophyll a concentration will appear in coastal areas in summer.Moreover,the influence of trade winds in the northern and southern hemispheres may be the reason of the upwelling,which leads to the increase of chlorophyll a concentration.