A spatiotemporal attention-augmented ConvLSTM model for ocean remote sensing reflectance prediction

• DOI: 10.1016/j.jag.2024.103815
• Publication Date: 2024-04-05T19:24:54Z
• Authors: Gaoxiang Zhou; Jun Chen; Ming Liu; Lingfei Ma
• Keywords: Environmental sciences; Physical geography; Ocean remote sensing; 0207 environmental engineering; Deep learning; GE1-350; 02 engineering and technology; Reflectance prediction; Spatiotemporal attention augmentation; GB3-5030

Remote sensing reflectance (Rrs) is an essential parameter in ocean color remote and a fundamental input for the estimation of elements. Predicting Rrs has potential to enable simultaneous prediction multiple marine environmental parameters, facilitating multi-perspective analysis changes. This paper proposes spatiotemporal attention-augmented ConvLSTM-based model prediction. The developed can predict up seven days by simultaneously learning features from time series auxiliary variables. According experiments, proposed achieves optimal performances on predictions at 443, 488, 555 nm, with Root Mean Squared Error (RMSE) Absolute Percentage (MAPE) first four less than 5.6*10-4 sr-1 8.6 %, respectively, which are better performance convolutional neural network (CNN), LSTM, CNN-LSTM, ConvLSTM. spatial temporal variations also compared evaluate effectiveness model, presenting consistent pattern between predicted observed Rrs. We found that integrating sea surface temperature (SST), photosynthetically available radiation (PAR), aerosol optical thickness 869 nm (AOT869) into improve accuracy various degrees. work suggests deep 7 convincing performance, providing critical data technical support ocean-related applications, such as algae bloom monitoring.