Oceanic internal wave amplitude retrieval from satellite images based on a data-driven transfer learning model

Internal waves (IW) are characterized by a large-amplitude, long-wave crest, and long-propagation distance. They are widespread in the global ocean. Amplitude is an essential IW parameter and is difficult to derive from the IW surface signatures in satellite images. A laboratory experiment and combined satellite/in-situ measurements were carried out to build two internal wave datasets (888 pairs of lab data and 121 pairs of synchronous in-situ data and satellite images). To efficiently use the lab data, we implemented a transfer learning model to retrieve IW amplitude from satellite images. The model is a purely data-driven model pre-trained with lab data and re-trained with satellite/in-situ data. A short connection was incorporated into the transfer learning framework to reduce information loss. Bias correction was adopted to improve the model performance. After the correction, the root mean square error (RMSE) of the estimated IW amplitude decreased from 12.09 m (17.84 m) to 9.59 m (11.59 m), the mean relative error decreased from 21% (27%) to 18% (16%), and the correlation coefficients improved from 0.81 (0.72) to 0.89 (0.90) on the test (training) dataset. For IWs with amplitude exceeding 100 m, the model can be expected to get an absolute error of 10 m. The mean relative error decreased with the increase in IW amplitudes. Comparisons with other algorithms demonstrate that the proposed model is efficient for IW studies. We applied the model to 156 satellite images containing IW signatures in the Andaman Sea, finding that large-amplitude IWs were mainly located at the water depth between 200 m and 1000 m on the continental slope. When considering one-pixel input errors for the peak-to-peak (PP) distance, the model shows large tolerance with the errors. Compared with the KdV equation-based method, the developed model was more accurate.