Remote Sensing Inversion Of Subsurface Chlorophyll Maxima Based On Convolutional Neural Network

Subsurface Chlorophyll Maxima(SCMs)are one of the significant commonalities of global marine biomass profiles,which refers to the peak chlorophyll phenomenon occurring in the water column deeper than the upper mixed layer,where the phytoplankton biomass and primary productivity contribute significantly to the whole water column.Since SCMs occur at depths deeper than the satellite detectable depth,there is still a lack of understanding of the large spatial and temporal scale distribution characteristics of this phenomenon.Making full use of the high-resolution ocean surface remote sensing data to realize the inversion of SCMs is a hot and difficult issue in statistical science and marine environmental science.In this paper,a SCM inversion model based on convolutional neural network structure is proposed for the first time,which is used to invert the vertical distribution of chlorophyll in the north-central South China Sea and analyze the spatial and temporal distribution of the depth and intensity characteristics of SCMs.In this paper,we first collected multi-source chlorophyll data from the South China Sea,including OCCCI water color remote sensing data,BGC-Argo buoy observation data and numerical results of 3D physical-ecological coupled model(CMEMS).Finally,the data are formatted and sliced into CMEMS model pre-training dataset and remote sensing-observation matched transfer learning dataset,and the training and test sets are divided respectively.Based on the convolutional neural network(CNN)model,a three-stage inversion model of vertical distribution of chlorophyll concentration(CNN-SCMs)is constructed in this paper,which are feature focusing stage,feature extraction stage and chlorophyll profile estimation stage.By comparing and analyzing the pre-training results of two different sets of input variables,it is determined that the CNN-SCMs model takes ocean surface chlorophyll concentration,sea surface temperature(SST)and its monthly characteristics as the input quantities,and the local area of the station where the chlorophyll profile observed by each BGC-Argo buoy is located as the input data range.In the feature focus phase,the model is focused on the central part of the local input of the ocean surface using the distance weight matrix to condense the input features of the model.In the feature extraction stage,the feature pyramid structure is used to extract the physical quantity features of the ocean surface at multiple spatial scales.In the chlorophyll profile estimation stage,the fully connected neural network structure and Sigmoid activation function are utilized to enhance the model's fit of the input features to the vertical chlorophyll profile.During the training process,the CNNSCMs model is first pre-trained using the CMEMS ocean model dataset,and the model parameters are tuned at this stage to obtain the model parameters of the optimal CNNSCMs and the initial weights of the model on the inverse chlorophyll vertical distribution task.The pre-trained CNN-SCMs are then applied to a dataset matching remote sensing data with BGC-Argo buoy observations for migration learning,finetuning the model using the observed data to more closely match the actual ocean physical phenomena,and finally validating and testing the model.In the model testing section,the accuracy of the CNN-SCMs model to invert SCMs is demonstrated in multiple dimensions from time and space.The pre-trained model has an MSE of 0.01 and a correlation coefficient r of 0.55 on the test set of ocean model data,and the results of the inversion of time series centered on SEATS station and the inversion of vertical distribution of chlorophyll in 116°E profile and 18°N profile are shown on the spatial and temporal scales,respectively,which demonstrate the effectiveness and accuracy of the trained model.After fine-tuning the model with remote sensing and observation data,the CNN-SCMs model has an MSE of 0.01 and a correlation coefficient of 0.84 in the test set.The results of comparing the vertical chlorophyll concentration distributions observed by two BGC-Argo buoys with the model inversion results at the spatial and temporal scales show that the CNNSCMs model can effectively invert the spatial and seasonal variations of the depth and intensity of SCMs in the north-central South China SeaFinally,using high-resolution remote sensing data from the central and northern South China Sea,the daily average high-resolution chlorophyll vertical distribution of the sea was obtained based on the inversion of the trained CNN-SCMs model,and the results of the long time series inversion of SCMs from 2014 to 2016 at SEATS stations were analyzed,which showed the yearly weakening of SCMs intensity,little change in SCMs depth and thickness,and significant seasonal changes.The seasonal variation is significant,the SCMs phenomenon disappears in winter,the SCMs depth is located between 50-80 m from spring to autumn,and the SCMs intensity ranges from 0.4 to 0.6 mg m-3.The mean vertical transects of 116°E and 18°N were analyzed for the long time series variation of SCMs and the spatial variation characteristics,and the results showed that the seasonal variation of the transects was consistent with the SEATS station,and the depth of SCMs was shallower and the concentration was lower in the nearshore area in space,while the SCM phenomenon was more obvious in the oceanic area,with the depth located between 60-80 m and the intensity range of 0.5-0.6 mg m-3.Finally,the spatial variation characteristics of the intensity and depth in the four seasons of 2020 within the study area in the South China Sea were analyzed,and the results showed that the depth and intensity of SCMs in the northern part of the South China Sea fluctuated more than those in the central part of the South China Sea,and the intensity of SCMs in the nearshore area was higher than that in the oceanic area.In summary,this paper proposes an inversion model of subsurface chlorophyll vertical distribution(CNN-SCMs)based on CNN neural network structure,introducing feature weight matrix,feature pyramid structure and global maximum pooling module to effectively fit the relationship between surface ocean physical quantities and subsurface chlorophyll vertical distribution in the north-central South China Sea as the study area.The results show that the model can effectively estimate the distribution of subsurface chlorophyll SCMs in the north-central South China Sea,which supports the construction of subsurface chlorophyll datasets in the South China Sea and the analysis of primary productivity studies in the South China Sea waters.