Deep Learning-based Inversion Study Of Chlorophyll Concentration

Chlorophyll-a(Chlor-a)is the main pigment in phytoplankton and also one of the indicators of marine biomass,how to accurately invert the global marine chlorophyll-a concentration is a worthy research topic.The Moderate Resolution Imaging Spectroradiometer(MODIS)is capable of collecting large amounts of water colour data.Meanwhile,deep learning,with its computational power and portability,has been extended to a number of fields,yet further work related to deep learning in the inversion of global marine chlorophyll a concentrations is urgently needed.In this paper,we use the Convolutional Neural Network(CNN)algorithm in deep learning to design the inversion model and a Recurrent Neural Network(RNN)algorithm was used to interpolate the inversion results,the main work and conclusions of this paper are as follows.(1)In this paper,we designed an inversion model of chlorophyll-a concentration based on convolutional neural network algorithm,and selected remote sensing reflectance(Rrs)images and chlorophyll a concentration images in five bands(412nm,469nm,488nm,547nm and 667nm)of MODIS-Aqua satellite based on chlorophyll a absorption spectra as experimental data,and performed data cleaning,region of interest selection and other pre-processing operations on the raw data.The hyperparameters such as convolution kernel and Patch were adjusted to design the CNN model to invert the model structure,and then generate the monthly average chlorophyll a concentration distribution for 12 months in2020.The mean coefficient of determination(R~2)was 0.930,the root mean squared error(RMSE)was 0.132 and the mean absolute error(MAE)was 0.103.The experimental results show that the CNN model designed in this paper can accurately invert the global ocean chlorophyll-a concentration.The experimental results show that the CNN model designed in this paper can accurately invert the global ocean chlorophyll-a concentration.(2)In this paper,a recurrent neural network based interpolation model for missing chlorophyll-a concentration data is designed to interpolate the missing data in the inversion results.The chlorophyll-a concentration images in the time range from 2005 to 2019 were selected as the experimental data.The Gated Recurrent Unit(GRU)model was preferred and further fine-tuned by comparing the experimental results of three RNN models.The results of the interpolation model trained with different number of fully connected neurons and network depth were compared to determine the final model parameters.The interpolation results of the GRU model were qualitatively and quantitatively analysed,and the R~2of the model was 0.958 and the RMSE was 0.087.The experimental results show that the GRU-based RNN model designed in this paper can effectively supplement the inversion missing data,improve the CNN model inversion distribution map,and increase the completeness of the inversion results.