Remote Sensing Retrieval Method Based on Few-Shot Learning: A Case Study of Surface Dissolved Organic Carbon in Jiangsu Coastal Waters, China

The ocean, Earth's largest carbon sink, plays a critical role in studying the carbon cycle in estuarine and coastal environments, making the retrieval of dissolved organic carbon (DOC) content highly significant. Machine learning retrieval techniques outperform traditional band ratio methods in handling high-dimensional and nonlinear features. However, the scarcity of in-situ observation data, a result of the challenges in acquiring marine environmental data, restricts the application of machine learning retrieval methods. To tackle this challenge, we propose a high-precision retrieval method for ocean surface DOC, combining spatial interpolation with deep neural networks (DNN), utilizing a small yet evenly distributed in-situ observation dataset. The same spatial interpolation technique was applied to both in-situ observation data and reflectance data of multiple bands at corresponding locations, expanding the original data distribution, thus increasing the sample size and diversity to create a virtual dataset. Subsequently, the machine learning model learned the complex relationships between target parameters and multidimensional features using the virtual dataset, which was then applied to real remote sensing reflectance images to obtain retrieval results. The validation results demonstrate that this method produces highly accurate DOC retrieval results (RMSE = 0.08 mg/L, MAPE = 4.13%, RPD = 3.28). This method fully leverages the limited in-situ observation data, offering a novel approach to quantitatively retrieving DOC concentrations via machine learning in the context of sparse observational data. This method holds potential for applications in large-scale ocean areas and for the retrieval of other dissolved substances.