Using Neural Network To Inverse The Chlorophyll-a In Non-cloud Missing Area In The Bohai Sea And Yellow Sea

Chla maps derived from satellite can provide large spatiotemporal coverage and near real-time synchronization observations of ocean color parameters.However,due to the presences of clouds,aerosols,sun glint and the incorrect masks in the high-turbidity coastal areas,lack of Chla retrievals are available,which represent an obstacle to observing ocean features and the spatiotemporal variations of Chla(non-cloud missing area).Selecting the Bohai Sea and the Yellow Sea as study area,based on the VIIRS scenes in 2017,a BP model is developed to inverse the Chla in the non-cloud missing area.The differences between the monthly Chl-BP derived from BP model and VIIRS monthly Chl-v,monthly valid pixels and monthly standard deviation are analysised.The main research contenes are as follow:Based on the 365 VIIRS scenes in 2017 and National Aeronautics and Space Administration(NASA)quality control flags to pick standard dataset.In order to reduce the influence of thin clouds and aerosols on Chla during the data processing,according to the spectral characteristics of Rayleigh-corrected reflectance(R_rc),the combination of the R_rc is used as the inputs of the model.Taking the complex composition of water and various kinds of aerosols into consideration,based on the theory of mapping,the Chla,K_d(490)and AOD(551)are as the outputs,which can enhance the stability of the model.The number of hidden layers and nodes of BP model are constantly adjusted,and the optimal BP architecture for this work is selected based on a maximal value of R² with a minimal number of neurons.Due to seasonal variations in aerosol and Chla,the monthly data were independently trained for modeling.The R²,RMSD,MAD,MRD and UPD of training are 0.94?0.99,0.33?1.56,14.19?36.61,5.86?24.01%and 5.69?15.39%,respensitively.The test results are the R² of 0.94?0.99,RMSD of 0.33?1.56,MAD of 14.19?36.61,MRD of 5.86?24.01%and UPD of 5.69?15.39%.The model is also applied to two independent satellite images,and the R² between the Chl-BP derived from model and VIIRS Chl-v was up to 0.98.The matching results with the measured data also show that the model has a good performance in recovering the Chla.The BP model was applied to the VIIRS 365 satellite sences in 2017 to count the monthly number of effective pixels.These results show that the valid pixels of Chl-BP are significantly increased.In highly turbid coastal waters,such as Bohai bay,Laizhou bay,liaozhou bay and Subei shoal,BP model can reconstruct the Chl-BP wrongly masked due to absorbing aerosol and high reflectance while ensuring the data quality.The monthly distributions of Chl-BP are consistent with the monthly Chl-v,and effectively fill the data gaps and increase the number of effective pixels of Chla,which is of great significance to long time series.The BP model can recover the data masked incorrectly due to thin clouds,aerosols and highly turbid coastal waters,and increase the effective coverage range of Chla,which is of profound significance to near-real time ocean dynamic changes and long time series.