Atmospheric Correction Over Oceans Based On Deep Learning

Satellite imagery is increasingly being used to monitor the oceanic environment.Atmospheric correction(AC)is essential before retrieving ocean color parameters from remote sensing data.This is because the water-leaving radiance only accounts for ～10%of the total signals received by the sensors.However,the current AC algorithms face several challenges,including the high uncertainty of remote sensing reflectance(Rrs),the difficulty in estimating the contributions of aerosol scattering,and the correction of cloud edge effects(CEEs).To address these problems,this paper proposes improved deep learning models for AC and the main results are as follows:1)Improvements of AC uncertainties in clear oceanic waters.While the goal of 5%for AC in the blue bands has been achieved,the uncertainties in the green and red bands are still high(over 10%).In this paper,a deep learning AC method(DLAC)was developed for oceanic waters.First,the standard near-infrared iterative AC(NIR)method was used to obtain the initial Rrs,and then the quality assurance(QA)system was employed to select high-quality remote sensing reflectance(Rrs)spectra.After that,the training dataset was randomly selected based on chlorophyll concentration intervals.Results show that DLAC learns the relationship between Rayleigh-corrected reflectance(Rrc)and Rrs very well.The band-averaged correlation coefficient on the validation dataset is 0.96,and the mean absolute percentage difference(MAPD)was less than 7%.In addition,the validation results using MOBY in situ measurements show that the performance of DLAC and NIR are closer in the blue and green bands.However,NIR shows many negative values in the red bands compared to DLAC,indicating that DLAC is more accurate in aerosol correction.Moreover,the uncertainties of the two AC methods were verified using an image-based chlorophyll constraint method.The results show that DLAC reduces the uncertainty by approximately 1%,5% and 15% in the blue,green and red bands respectively.DLAC can significantly improve the overor underestimation of pixels around stray light/sun glint.These benefits help to reveal the realistic patterns of oceanic waters.2)AC for turbid coastal waters.In the absence of shortwave infrared bands(e.g.GOCI),it is often difficult to realize AC for turbid coastal waters.This paper proposed a neural network-based AC model(NN＿3S)for coastal waters that did not rely on the shortwave infrared band.Similar to DLAC,NN＿3S combined the spectral quality assessment method with the deep learning model to build the AC algorithm.The difference was that NN＿3S adopted the optical classification and spectral scoring system(OC＿3S)proposed in this paper,which was based on the global hyperspectral datasets and was more suitable for turbid coastal waters.The results of the validation using AERONET-OC measurements show that NN＿3S achieves a better performance than the NIR and GOCI standard atmospheric correction(KOSC)methods,with a 17.4%and 32.2%reduction in MAPD compared to the two methods,respectively.Moreover,NN＿3S performs more consistently for different water types.The spatial distribution of ocean color parameters obtained by NN＿3S is consistent with the results of NIR and KOSC,but NN＿3S is able to obtain more valid data in turbid waters,thus helping to solve the problem of insufficient observations in turbid areas.The daily percentage of valid observations(DPVO)shows that NN＿3S is able to increase DPVO by 12%and102%in June and 32%and 86%in December compared to NIR and KOSC.NN＿3S provides an AC method over coastal waters for sensors without a shortwave infrared band.3)Cloud edge effect(CEE)correction.CEEs,such as cloud shadows,adjacent effects(AEs)and stray lights,severely degrade the quality of remote sensing images and greatly reduce the number of valid observations.In this paper,a deep learning technique for CEEs correction(DLTCC)is developed using the high-frequency observations of GOCI.First,the variation of seawater at different times was examined.It was found that the variation of water constituents could be largely maintained when the time interval was 1 hour.The Rrs from cloud-free observations at noon(11:55 and12:55 local time)of GOCI were then used as the true values to match the Rrc values affected by CEEs with the time difference of 1 hour.DLTCC was then used to learn the correction mechanism of CEEs.The comparison results based on AERONET-OC observations show that DLTCC achieves a comparable accuracy to that of NIR and KOSC with more valid data,demonstrating the practicality of the theoretical design of DLTCC.Secondly,the spatial analysis shows that DLTCC significantly reduces the impact of CEEs.For example,in the blue bands,the CEEs can be reduced by about 10%for two pixels from the cloud edges,while in the green band it can reach 50%.Finally,DLTCC substantially increases the number of valid observations.Compared to NIR and KOSC,the DPVO of DLTCC increases by more than 71%,which helps to accurately reveal the spatio-temporal variation of bio-optical parameters.